Pith. sign in

REVIEW 3 major objections 6 minor 34 references

Benchmarking and Reasoning Distillation of Large Language Models for Feedback Controller Design in Complex Dynamical Systems

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a 1.5B-parameter model trained on a stronger model's reasoning traces can design feedback controllers for complex dynamical systems well enough to run on edge hardware and track targets on a real…

desk verdict A genuine benchmark extension and a plausible distillation result, but the absolute claims need a classical baseline and artifact release before they become load-bearing. read the letter →

arxiv 2608.07004 v1 pith:BPZGF53G submitted 2026-08-07 cs.RO cs.SYeess.SY

Zhongchao Zhou , Yixuan Xie , Wenwei Yu , Yuxi Lu , Yaonan Zhu , Qian Niu , Yutaka Matsuo , Yusuke Iwasawa This is my paper · ORCID
classification cs.ROcs.SYeess.SY
keywords LLM-basedcontrollerdesignfeedbackcontrolreasoningdistillationedgedeploymentmulti-DoFdynamicalsystemsPIDtuningslidingmodepneumaticartificialmuscle
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 sets out to test whether large language models can design feedback controllers for complex dynamical systems—not just linear single-degree-of-freedom textbook plants—and whether a small model can do the job on edge hardware. To that end it introduces CoDyControlBench, 132 system configurations spanning 1-6 degrees of freedom, four system types, two coupling levels, three damping regimes, and two controller families. Six LLMs are scored on executable-code controller design; the strongest model reaches a 94.8% design success rate and the weakest reaches 50.0%. The central constructive claim is that reasoning distillation, training a 1.5B student on the teacher's diagnostic reasoning traces rather than only on final answers, produces a model that stays between roughly 54% and 69% success across all DoF levels and succeeds in all three physical trials on a pneumatic-artificial-muscle robotic arm. If true, lightweight, locally executed controller-design assistants are feasible without cloud-hosted frontier models.

What carries the argument

The load-bearing artifact is CoDyControlBench itself: 132 parameterized mass-spring-damper systems in state-space form, each task requiring executable PID-family or sliding-mode-family controller code filled into a fixed template. Performance is first gated by pass/fail criteria—all outputs inside a ±2% steady-state band for the final second and overshoot below 100%—and then scored by Equation (7), a weighted penalty on steady-state error, overshoot, delay, and terminal oscillation. The second mechanism is reasoning distillation: the same 599 LTI tuning episodes train a 1.5B DeepSeek student either with teacher-generated diagnostic traces (Think-Model) or with concise rationales and code only (Answer-Model), with the small-model evaluation relaxed to a ±5% band and ten tuning iterations. The benchmark protocol plus the reasoning-trace supervision together carry the argument.

What would settle it

Run the same six models and the Think-Model on CoDyControlBench with the steady-state band tightened from ±2% to ±0.5% and the overshoot cap lowered to 20%, or inject a disturbance mid-simulation; if GPT's 94.8% success or the Think-Model's 54-69% stability collapses or reorders, the selected thresholds were carrying the result.

Watch

Extended reading notes

Core claim

The paper's central discovery is twofold. First, on CoDyControlBench a frontier commercial model (GPT) is the only evaluated model whose design success stays above 88% from 1 to 6 DoFs, while the largest aggregate performance swings come from number of DoFs and controller family rather than system type, coupling, or damping. Second, a 1.5B student distilled on GPT's reasoning traces reaches 54-69% success across all DoF levels, outperforms the answer-distilled variant, and achieves successful tracking in 3/3 physical trials on a PAM-driven arm. The paper attributes the best-versus-worst model gap to control-design knowledge—gain scaling and transient-limiting mechanisms such as output saturation and anti-windup—rather than to failures in generating executable code, since code-extraction failures are rare for both models.

Load-bearing premise

The results rest on the assumption that a trial passing the benchmark's hand-set thresholds—±2% steady-state band over the last second and overshoot under 100%, relaxed to ±5% and ten iterations for the small models—is a faithful proxy for successful controller design; if those thresholds are too loose or too task-specific, the absolute success rates and the edge-deployment conclusion do not transfer to real systems, even if relative model comparisons survive.

Editorial extensions

If this is right

  • Edge deployment becomes practical: a 1.5B model running locally on a Jetson-class device can iteratively tune a PID controller to track a target on a nonlinear, hysteretic PAM arm.
  • DoF count and controller type are the benchmark dimensions that most separate models, so future controller-LLM work should emphasize high-dimensional dynamics and heuristic PID tuning rather than coupling or damping effects.
  • LLMs write executable sliding-mode controllers more reliably than they search PID gains, yet successful PID designs score better, revealing a trade-off between design feasibility and control quality.
  • The dominant failure mode of weaker models is mis-scaled gains, not non-executable code: the weakest model's controllers end in steady-state deviation above tolerance in 68.3% of attempts while diverging in only 0.2%.
  • Reasoning distillation transfers response-driven tuning strategies across system dimensions, while answer-only distillation does not generalize beyond low-DoF systems.

Reading between the lines

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

  • Testable extension: tighten the benchmark gate from ±2% to ±0.5% steady-state band and lower the overshoot cap from 100% to 20%; the absolute success rates and possibly the model rankings would shift, testing whether the thresholds were carrying the result.
  • The distillation was trained only on LTI plants, so its success on nonlinear and time-varying benchmark cases suggests the student learned generic response-driven heuristics rather than memorized gains; a direct ablation would distill from a weaker teacher or shorter traces and measure how much of the Think-Model's margin survives.
  • Even the Think-Model fails in roughly 30-45% of benchmark trials, so a real edge deployment would likely need a fallback controller or human check rather than fully autonomous closed-loop authority.
  • The physical PAM validation used three trials at one target angle; scaling to more targets and adding disturbances would test whether the 3/3 success reflects generalizable tuning skill or favorable test conditions.
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.

Referee Report

3 major / 6 minor

Summary. The paper introduces CoDyControlBench, a benchmark of 132 system configurations (spanning 1–6 DoFs, four system types, two coupling levels, three damping regimes, and two controller families) for evaluating LLM-based feedback controller design, and reports a three-run evaluation of six LLMs (GPT, Gemini, Claude, GLM, DeepSeek, Qwen). The authors find that GPT achieves the highest success rate (94.8%) and Qwen the lowest (50.0%), identify number of DoFs and controller type as the largest sources of variation, and compare GPT with Qwen to attribute the gap to gain selection and transient-limiting mechanisms. They then distill GPT-5.5 reasoning traces into a 1.5B-parameter model (Think-Model) using matched answer-distillation and base-model controls, report that the Think-Model maintains roughly stable success rates (54–69%) across 1–6 DoFs on a relaxed evaluation protocol, and validate the Think-Model on a one-DoF PAM-driven robotic arm where it achieves three-out-of-three successful tracking trials.

Significance. If the results hold, the paper provides a useful new benchmark for a relatively underexplored application of LLMs, and the matched answer-vs-reasoning distillation comparison is a valuable contribution: the Think-Model's edge over the Answer-Model and the base model is internally consistent, and the real-world PAM trials give a concrete demonstration of a 1.5B controller-design model on edge hardware. The paper is also transparent about several limitations, including the statement in Section II-D that the pass/fail criteria are for 'benchmark-level feasibility screening rather than formal safety verification' and the explicit disclosure of the relaxed protocol for the small models in Section III-C. However, the benchmark's validity as a measure of controller-design ability is not yet established, because no non-LLM control baseline is reported; the absolute success rates and the edge-deployment feasibility claim depend on the uncalibrated pass/fail threshold, and the small-model performance claims are framed in the abstract without the protocol caveats. These issues are addressable and do not undermine the relative comparisons within the paper.

major comments (3)
  1. [Section II-D, Eq. (7), pass/fail criteria] The evaluation protocol is uncalibrated because no non-LLM control baseline is reported. The manuscript itself states that the pass/fail criteria are 'designed for benchmark-level feasibility screening rather than formal safety verification,' and the criteria (all final-1-s samples within a ±2% band and overshoot below 100%) are loose enough that a scripted high-gain saturated PID with anti-windup could plausibly pass most of the 132 fully actuated, constant-input-matrix plants. Section V provides supporting evidence: GPT's advantage over Qwen comes largely from aggressive gains plus output saturation (98% vs 21% use of saturation), which is exactly the behavior a lenient pass band rewards. Without a classical or scripted baseline (e.g., Ziegler-Nichols-tuned PID, a manually saturated high-gain PID, or an LQR) evaluated under the same protocol, the absolute success rates (94.8%, 50.0%, Think-Model ~60%) do not establish LLM controller-design capability, and the edge-deployment feasibility claim rests on an uncalibrated metric. Please add such a baseline and report its success rate and score.
  2. [Section III-C and Table VIII] The evaluation protocol for the 1.5B models and DeepSeek-V4-Pro differs from the main benchmark: the steady-state band is relaxed from 2% to 5%, fixed iterations increase from 3 to 10, and the text indicates only PID-family tasks were used in this comparison. The abstract and conclusion state that the Think-Model 'maintained stable performance across 1–6 DoFs' and 'outperformed the answer-distilled and base model on CoDyControlBench' without disclosing these relaxations. Because Table VIII's success rates are not directly comparable to Table I, the reader may overinterpret the distillation result. Please either evaluate the small models under the original 2%/Fixed-3 protocol as well, or explicitly state in the abstract and conclusion that the small-model claims are based on the relaxed protocol and the PID-only subset.
  3. [Sections II.A–II.B and Appendix A] The composition of the 132 configurations is under-specified. The paper does not provide a table giving the number of systems per factor level (e.g., counts for each DoF by system type, damping regime, and coupling level), nor the total number of design tasks (each configuration appears to have both PID and SMC tasks, implying 264 tasks). The benchmark prompts and evaluation code are referenced as 'benchmark files' but no repository URL or supplementary file is included in the manuscript. Without this information, the benchmark cannot be reproduced, and the factor-level ranges in Table VII cannot be interpreted. Please include the full configuration grid and a release link.
minor comments (6)
  1. [Abstract] The abstract contains a typo: 'traget tracking' should be 'target tracking.'
  2. [Section III.C] The sentence 'For a fair comparison, DeepSeek-V4-Pro is additionally evaluated under the same configuration (PID, Fixed-10, 5% steady-state band)' is ambiguous if the 1.5B models were also evaluated on PID-only tasks; please state explicitly the controller-type scope of Table VIII.
  3. [Section II.D] The score weights (w_ess=2, w_mp=2, w_td=1) are hand-set; a brief sensitivity analysis (e.g., varying the weights and checking whether the score-based conclusions in Section IV change) would strengthen the control-quality comparisons.
  4. [Section V] The phrase '3DoF levels×24systems per level' should read '3 DoF levels (DoF=4,5,6) × 24 systems per level' for clarity.
  5. [Section IV.B] The phrase 'Compare to it' should be 'Compared to it.'
  6. [Section III.C / Fig. 4] The PAM experiment section would benefit from stating the quantitative success criteria (5% steady-state band over the final time window) before reporting the 100% vs 33.3% success rates.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmark results and distillation comparisons are empirical and self-contained.

full rationale

The paper's central claims are empirical benchmark results and physical experiments rather than a derivation chain, so there is no equation-level circularity. The scoring rule (Eq. 7) is a hand-specified evaluation metric with stated weights (w_ess=2, w_mp=2, w_td=1); it is not fitted to any model output and no predicted quantity is reused as an input. The relaxation of the pass/fail band from 2% to 5% and the iteration limit from 3 to 10 for the 1.5B models (Section III-C) is applied uniformly to the base model, Answer-Model, Think-Model, and DeepSeek-V4-Pro, so the Think-Model's advantage is not forced by construction. The paper explicitly labels its criteria as benchmark-level feasibility screening rather than formal safety verification (Section II-D), which is a validity caveat, not a circular step. Self-citations such as [9], [22], [27], [31], and [34] are used for background, motivation, or hardware descriptions; none is invoked as a uniqueness theorem or as the sole justification of the benchmark's conclusions. In particular, the PAM hardware reference [34] supports the physical setup, while the successful tracking results are newly measured in this work. No fitted parameter is renamed as a prediction, no known result is merely relabeled, and no ansatz is imported solely through a self-citation. The relative model rankings and the physical trials have independent content, so the appropriate finding is no significant circularity.

Assumptions & free parameters 7 free parameters · 6 assumptions · 0 invented entities

The central claims rest on hand-set evaluation thresholds, protocol choices, and unstated distributional assumptions. There are no fitted physical constants or newly invented physical entities; the free parameters are all in the evaluation and training protocol.

free parameters (7)
  • Score weights (w_ess, w_mp, w_td) = 2, 2, 1
    Eq. (7): set 'empirically' to penalize steady-state error and overshoot more than delay; no sensitivity analysis is provided.
  • Steady-state pass/fail band = 2% (main models), 5% (distilled models)
    Section II-D and III-C: the band defines success; it was relaxed for 1.5B models to avoid floor effects, directly affecting success-rate magnitudes.
  • Maximum overshoot threshold = 100% of target
    Section II-D: a run fails if overshoot exceeds 100%; this threshold is adjustable and strongly influences which controllers count as successful.
  • Terminal oscillation penalty parameters = P_osc=5; >=3 directional reversals; >=0.5% peak-to-peak; ignore <0.1% changes
    Section II-D: ad hoc oscillation detector used in the score; affects ranking of controllers that converge but oscillate.
  • Maximum tuning iterations = 3 (Fixed-3 / Success-Stop), 10 (distilled models)
    Protocol choice in Section II-B and III-C; more iterations make design easier, so the absolute success rates are protocol-dependent.
  • Damping regime boundaries = overdamped zeta>=2; critical zeta in [0.9,1.1]; underdamped zeta<0.1
    Section II-A: categorical boundaries chosen for the benchmark; results across regimes depend on these cutoffs.
  • Distillation training hyperparameters = QLoRA rank 16, alpha 32, dropout 0.05, LR 1.5e-4, batch 8, 3 epochs
    Section III-B: chosen by the authors; the distilled model's performance could vary with these settings.
assumptions (6)
  • domain assumption Plants are fully actuated second-order systems with x_dot = A(x,t)x + Bu and constant input matrix B
    Section II-A: all benchmark systems share this structure; results may not generalize to underactuated or B(x,t) systems.
  • domain assumption The pass/fail criteria and the score in Eq. (7) are valid measures of controller quality
    Section II-D: success is defined by the authors' thresholds, not by an external standard or human expert labels.
  • domain assumption Simulation-based evaluation with the provided code template reflects real controller-design capability
    Section II-B: the LLM only fills [TODOs] in a fixed template; the benchmark measures parameter and law selection under a known plant model, not full controller design from scratch.
  • ad hoc to paper The 599 LTI training episodes are sufficiently out-of-distribution from CoDyControlBench to test generalization
    Section III-B: the paper states parameters were sampled from separately defined distributions but does not specify them or quantify overlap.
  • ad hoc to paper GPT-5.5 reasoning traces are correct and complete enough to serve as supervision for the distilled model
    Section III-A: no independent verification of the reasoning traces is provided.
  • domain assumption Three independent runs and three physical trials provide stable estimates
    Section II-D and III-C: standard deviations are large in several cells, and the physical validation has n=3 on one robot.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Benchmarking and Reasoning Distillation of Large Language Models for Feedback Controller Design in Complex Dynamical Systems." pith.science (2026). https://pith.science/paper/BPZGF53G

@misc{pith2026260807004,
  author       = {Pith},
  title        = {Pith review of: Benchmarking and Reasoning Distillation of Large Language Models for Feedback Controller Design in Complex Dynamical Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BPZGF53G}},
  note         = {Machine review of arXiv:2608.07004}
}
read the original abstract

Although remarkable capabilities have been demonstrated by Large Language Models (LLMs) across scientific domains, feedback controller design remains underexplored. Existing benchmarks focus mainly on linear single-Degree-of-Freedom (DoF) systems and large API-hosted models, leaving performance on complex controller-design tasks and feasibility for edge deployment unclear. To address these limitations, we introduce the Complex Dynamics-to-Control Benchmark for Large Language Models (CoDyControlBench), comprising 132 system configurations across five evaluation dimensions: number of DoF, system type, coupling level, damping regime, and controller type. Six state-of-the-art LLMs were evaluated over three independent runs, including three commercial models (GPT, Gemini, and Claude) and three open-source models (GLM, DeepSeek, and Qwen). GPT achieved the highest design success rate at 94.8\%, whereas Qwen showed the lowest rate at 50.0\%. Across the benchmark dimensions, DoF and controller type exhibited the largest model-averaged variations in design success, with success-rate ranges of 36.3\% and 17.6\%, respectively, exceeding those associated with system type, coupling level, and damping regime. Comparison of GPT and Qwen showed that their performance gap arose mainly from the control-design knowledge, particularly gain selection and the use of transient-limiting mechanisms. For edge deployment, a specialized 1.5B-parameter model was developed through reasoning distillation. The reasoning-distilled model outperformed the answer-distilled and base model on CoDyControlBench, maintained stable performance across 1-6 DoFs, and achieved successful traget tracking in all three physical trials on a pneumatic-artificial-muscle-driven robotic arm. These results establish a benchmark baseline and highlight the potential of lightweight, edge-deployable controller-design models.

Figures

Figures reproduced from arXiv: 2608.07004 by the authors.

Figure 1
Figure 1. Design principle of CoDyControlBench • Fully independent: Each DoF is dynamically independent, allowing the system to be decomposed into separable single-input single-output subsystems. • Fully coupled: Neighboring DoFs are dynamically cou￾pled through shared states or torque interactions, resulting in multi-input multi-output dynamics. Generally, stronger system coupling can increase controller￾design complexity [2… view at source ↗
Figure 2
Figure 2. Illustration of Q60. B =             0 0 0 1 m1 0 0 0 0 0 0 1 m2 0 0 0 0 0 0 1 m3             . (5) The masses and generalized coordinates are m1 = m2 = m3 = 1, q1 = x1, q2 = x3, q3 = x5. The time-varying base stiffness coefficients are k11(t) = 3.0 + 0.4 sin t, k22(t) = 2.6 + 0.3 sin(0.8t), k33(t) = 2.8 + 0.35 cos(0.6t), k12(t) = k21(t) = 1.0 + 0.2 cost, k13(t) = k31(t) = 0.8 + 0.15 sin(0.5t… view at source ↗
Figure 3
Figure 3. Process of Performance Assessment. D. Benchmark Assessment and Model Selection Evaluating controller performance is inherently challenging, as it depends on multiple dimensions of system behavior. To ensure a consistent and comparable assessment across different control tasks, this benchmark adopts several widely used metrics that jointly reflect steady-state accuracy, transient characteristics, and response speed. … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Control responses of (a) the Think-Model and (b) the Answer-Model [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 27 canonical work pages

  1. [1]

    A shared control framework for enhanced grasping performance in teleoperation,

    Y . Zhu, B. Jiang, Q. Chen, T. Aoyama, and Y . Hasegawa, “A shared control framework for enhanced grasping performance in teleoperation,” IEEE Access, vol. 11, pp. 69 204–69 215, 2023

    work page 2023

  2. [2]

    Uncertainty- aware hybrid modeling for spatial thermal forecasting in historic buildings with incomplete structural information,

    F. Kang, H. Lin, Z. Zhou, Y . Ren, X. Li, and J. Wang, “Uncertainty- aware hybrid modeling for spatial thermal forecasting in historic buildings with incomplete structural information,”Journal of Building Engineering, p. 115231, 2026

    work page 2026

  3. [3]

    Comparison study between open-loop and closed-loop identification for industrial hydraulics actuator system,

    N. H. M. Ali, R. Ghazali, H. I. Jaafar, M. F. Ghani, C. C. Soon, and Z. Has, “Comparison study between open-loop and closed-loop identification for industrial hydraulics actuator system,”International Journal of Mechanical Engineering and Robotics Research, vol. 13, no. 5, pp. 516–521, 2024

    work page 2024

  4. [4]

    Ogata,Modern control engineering

    K. Ogata,Modern control engineering. Prentice hall, 2010

    work page 2010

  5. [5]

    A bridge between state-space and transfer-function meth- ods,

    V . Ku ˇcera, “A bridge between state-space and transfer-function meth- ods,”Annual Reviews in Control, vol. 23, pp. 177–184, 1999

    work page 1999

  6. [6]

    Data-driven methods applied to soft robot modeling and control: A review,

    Z. Chen, F. Renda, A. Le Gall, L. Mocellin, M. Bernabei, T. Dangel, G. Ciuti, M. Cianchetti, and C. Stefanini, “Data-driven methods applied to soft robot modeling and control: A review,”IEEE Transactions on Automation Science and Engineering, vol. 22, pp. 2241–2256, 2024

    work page 2024

  7. [7]

    Robust safety-critical control for dynamic robotics,

    Q. Nguyen and K. Sreenath, “Robust safety-critical control for dynamic robotics,”IEEE Transactions on Automatic Control, vol. 67, no. 3, pp. 1073–1088, 2021

    work page 2021

  8. [8]

    Llm- based test-driven interactive code generation: User study and empirical evaluation,

    S. Fakhoury, A. Naik, G. Sakkas, S. Chakraborty, and S. K. Lahiri, “Llm- based test-driven interactive code generation: User study and empirical evaluation,”IEEE Transactions on Software Engineering, 2024

    work page 2024

Show all 34 references
  1. [9]

    Large lan- guage models are zero-shot reasoners,

    T. Kojima, S. S. Gu, M. Reid, Y . Matsuo, and Y . Iwasawa, “Large lan- guage models are zero-shot reasoners,”Advances in neural information processing systems, vol. 35, pp. 22 199–22 213, 2022

    2022

  2. [10]

    Mathematical capabilities of chatgpt,

    S. Frieder, L. Pinchetti, R.-R. Griffiths, T. Salvatori, T. Lukasiewicz, P. Petersen, and J. Berner, “Mathematical capabilities of chatgpt,” Advances in neural information processing systems, vol. 36, pp. 27 699– 27 744, 2023

    2023

  3. [11]

    Toolllm: Facilitating large language models to master 16000+ real-world apis,

    Y . Qin, S. Liang, Y . Ye, K. Zhu, L. Yan, Y . Lu, Y . Lin, X. Cong, X. Tang, B. Qianet al., “Toolllm: Facilitating large language models to master 16000+ real-world apis,”arXiv preprint arXiv:2307.16789, 2023

    2023 arXiv

  4. [12]

    Llm adaptive pid control for b5g truck platooning systems,

    I. de Zarz `a, J. de Curt `o, G. Roig, and C. T. Calafate, “Llm adaptive pid control for b5g truck platooning systems,”Sensors, vol. 23, no. 13, p. 5899, 2023

    2023

  5. [13]

    Controlagent: Automating control system design via novel integration of llm agents and domain expertise,

    X. Guo, D. Keivan, U. Syed, L. Qin, H. Zhang, G. Dullerud, P. Seiler, and B. Hu, “Controlagent: Automating control system design via novel integration of llm agents and domain expertise,”arXiv preprint arXiv:2410.19811, 2024

    2024 arXiv

  6. [14]

    Agenticcontrol: An automated con- trol design framework using large language models,

    M. Narimani and S. A. Emami, “Agenticcontrol: An automated con- trol design framework using large language models,”arXiv preprint arXiv:2506.19160, 2025

    2025 arXiv

  7. [15]

    Smartcontrol: Interactive pid controller design powered by llm agents and control system expertise,

    K. Tohma, H. ˙I. Okur, H. G ¨ursoy-Demir, M. N. Aydın, and C. Yero ˘glu, “Smartcontrol: Interactive pid controller design powered by llm agents and control system expertise,”SoftwareX, vol. 31, p. 102194, 2025. IEEE TRANSACTIONS ON XXX, VOL. XX, NO. X, 2026 11

    2025

  8. [16]

    Large language model- based tuning assistant for variable speed pmsm drive with cascade control structure,

    T. Tarczewski, D. Stojic, and A. Dzielinski, “Large language model- based tuning assistant for variable speed pmsm drive with cascade control structure,”Electronics, vol. 14, no. 2, p. 232, 2025

    2025

  9. [17]

    Agentic ai for real-time adaptive pid control of servo motor,

    T. M. Arif and M. A. Rahim, “Agentic ai for real-time adaptive pid control of servo motor,” inActuators, vol. 14, no. 9. MDPI AG, 2025, p. 459

    2025

  10. [18]

    Toward engineering agi: Benchmarking the en- gineering design capabilities of llms,

    X. Guo, Y . Li, X. Kong, Y . Jiang, X. Zhao, Z. Gong, Y . Zhang, D. Li, T. Sang, B. Zhuet al., “Toward engineering agi: Benchmarking the en- gineering design capabilities of llms,”arXiv preprint arXiv:2509.16204, 2025

    2025

  11. [19]

    Llm-based pid controller optimization,

    I. Kamenko, S. Ilic, and V . Congradac, “Llm-based pid controller optimization,” in2025 10th International Conference on Smart and Sustainable Technologies (SpliTech). IEEE, 2025, pp. 1–6

    2025

  12. [20]

    Llm- agent-controller: A universal multi-agent large language model system as a control engineer,

    R. Zahedifar, S. A. Mirghasemi, M. S. Baghshah, and A. Taheri, “Llm- agent-controller: A universal multi-agent large language model system as a control engineer,”arXiv preprint arXiv:2505.19567, 2025

    2025 arXiv

  13. [21]

    Mrac-llm toolbox: An interactive model reference adaptive control enhanced with large language models,

    M. N. Aydın, H. I. Okur, H. G ¨ursoy-Demir, K. Tohma, and C. Yero ˘glu, “Mrac-llm toolbox: An interactive model reference adaptive control enhanced with large language models,”SoftwareX, vol. 34, p. 102556, 2026

    2026

  14. [22]

    Llm-guided adaptive compensator: Bringing adaptivity to robotic feedback control with large language model,

    Z. Zhou, Y . Lu, Y . Zhu, Y . Zhao, Q. Niu, B. He, L. He, W. Yu, and Y . Iwasawa, “Llm-guided adaptive compensator: Bringing adaptivity to robotic feedback control with large language model,”IEEE Transactions on Automation Science and Engineering, 2026

    2026

  15. [23]

    When control meets large language models: From words to dynamics,

    K. Nosrati, A. Tepljakov, J. Belikov, and E. Petlenkov, “When control meets large language models: From words to dynamics,”Engineering Applications of Artificial Intelligence, vol. 178, no. 2, p. 115119, 2026

    2026

  16. [24]

    Capabilities of large language models in control engineering: A benchmark study on gpt-4, claude 3 opus, and gemini 1.0 ultra,

    D. Kevian, U. Syed, X. Guo, A. Havens, G. Dullerud, P. Seiler, L. Qin, and B. Hu, “Capabilities of large language models in control engineering: A benchmark study on gpt-4, claude 3 opus, and gemini 1.0 ultra,”arXiv preprint arXiv:2404.03647, 2024

    2024 arXiv

  17. [25]

    J. J. DiStefano, A. J. Stubberud, and I. J. Williams,Schaum’s outline of feedback and control systems. McGraw-Hill Professional, 1997

    1997

  18. [26]

    N. S. Nise,Control systems engineering. John Wiley & Sons, 2019

    2019

  19. [27]

    A gan based pid controller for highly adaptive control of a pneumatic-artificial-muscle driven antagonistic joint,

    Z. Zhou, Y . Lu, S. Kokubu, P. E. Tort ´os, and W. Yu, “A gan based pid controller for highly adaptive control of a pneumatic-artificial-muscle driven antagonistic joint,”Complex & Intelligent Systems, vol. 10, no. 5, pp. 6231–6248, 2024

    2024

  20. [28]

    M. W. Spong, S. Hutchinson, and M. Vidyasagar,Robot Modeling and Control. John Wiley & Sons, 2005

    2005

  21. [29]

    Skogestad and I

    S. Skogestad and I. Postlethwaite,Multivariable Feedback Control: Analysis and Design. John Wiley & Sons, 2005, vol. 2

    2005

  22. [30]

    Model- based control of soft robots: A survey of the state of the art and open challenges,

    C. Della Santina, R. K. Katzschmann, T. M. Bieze, and D. Rus, “Model- based control of soft robots: A survey of the state of the art and open challenges,”Frontiers in Robotics and AI, vol. 8, p. 639167, 2021

    2021

  23. [31]

    Offline-trained gan-augmented highly adaptive control with multi-dof fusion for pneumatic soft surgical robots,

    Y . Lu, Z. Zhou, D. Zheng, Y . Zhou, Z. Wang, S. Jiang, W. Yu, and B. He, “Offline-trained gan-augmented highly adaptive control with multi-dof fusion for pneumatic soft surgical robots,”IEEE Transactions on Automation Science and Engineering, vol. 23, pp. 942–954, 2025

    2025

  24. [32]

    Experimental comparison of control strategies for a periph- eral nerve interface actuator,

    J. J. Garc ´ıa-Valdovinos, T. Salgado-Jim ´enez, H. V . R´ıos-Figueroa, and I. Chairez, “Experimental comparison of control strategies for a periph- eral nerve interface actuator,”Frontiers in Robotics and AI, vol. 6, p. 11, 2019

    2019

  25. [33]

    Cloud to edge: Bench- marking llm inference on hardware-accelerated single-board computers,

    H. Renney, F. Trad, M. Mattarock, and Z. Wood, “Cloud to edge: Bench- marking llm inference on hardware-accelerated single-board computers,” arXiv preprint arXiv:2604.24785, 2026

    2026 arXiv

  26. [34]

    Optimization of spring constant of a pneumatic artificial muscle-spring driven antagonistic structure,

    Z. Zhou, S. Kokubu, Y . Wang, Y . Lu, P. E. Tort ´os, and W. Yu, “Optimization of spring constant of a pneumatic artificial muscle-spring driven antagonistic structure,”IEEE Robotics and Automation Letters, vol. 7, no. 3, pp. 5982–5989, 2022

    2022

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.