Pith. sign in

REVIEW 4 major objections 6 minor 4 cited by

Autonomous Control Leveraging LLMs: An Agentic Framework for Next-Generation Industrial Automation

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read LLM agents whose proposals are checked and reprompted on rejection achieve perfect fault recovery in finite-state machines and near-PID temperature control on a physical rig.

desk verdict Solid FSM planning results and an honest failure-mode analysis, but the continuous-control claims outrun the evidence: the digital twin is unvalidated and the abstract promises an ablation that isn't there. read the letter →

arxiv 2507.07115 v1 pith:MIE4N62L submitted 2025-07-03 cs.AI cs.MAcs.SYeess.SY

Javal Vyas , Mehmet Mercangoz This is my paper
classification cs.AIcs.MAcs.SYeess.SY
keywords AutonomousControlLargeLanguageModelsAgenticAIFrameworkIndustrialAutomationZero-ShotLearningFiniteStateMachinesDigitalTwinValidator-RepromptingLoop
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 tries to establish that large language models, which are unreliable at multi-step reasoning on their own, become dependable industrial-automation components when wrapped in a closed loop that checks every proposal and reprompts on rejection. In the discrete case, an LLM proposes recovery routes through a finite-state machine, a simulator walks the route, a validator rejects invalid transitions, and a reprompting agent generates a corrected prompt; with this loop, GPT-4o and GPT-4o-mini produce a valid recovery path on all 180 random state machines within five attempts, staying near the shortest path. In the continuous case, the same pattern of proposing heater powers, validating predicted temperature against a digital twin, and reprompting on violation regulates a dual-heater lab system under a persistent cooling disturbance with accuracy the authors characterize as approaching a classical PID baseline. The significance is that a single language-driven architecture could span the high-level symbolic reasoning and low-level continuous control that industrial automation currently splits across rigid rule-based systems, opening a path toward systems that can improvise responses to rare faults rather than only execute pre-programmed ones.

What carries the argument

The central mechanism is the validator-reprompting loop: a cycle in which a proposed action is executed in a low-cost checker (the FSM dictionary or a digital-twin simulator), a validation agent applies rule-based criteria (transition validity, power bounds of 0 to 0.3 W, progress of the average temperature toward setpoint), and on rejection a reprompting agent rewrites the action agent's prompt with the specific violation highlighted before the next proposal. The loop is bounded by a five-iteration budget in the FSM study and by a fixed number of inference cycles in the temperature study, after which a safety override hands control to a conservative fallback policy. Two further objects carry the argument: finite state machines used as interpretable operating envelopes, encoded as Python dictionaries that the authors find LLMs process more reliably than adjacency matrices or edge lists, and a digital twin of the TCLab heaters with fixed thermal parameters that makes validation deterministic and decouples controller evaluation from physical variability. A fixed 30-second planning interval absorbs LLM inference latency so the loop can act as a real-time controller.

What would settle it

Run step-response trials on the physical TCLab under the fan disturbance and compare measured temperatures against the digital twin's fixed-parameter predictions over 30-second intervals; if the twin frequently certifies actions that the physical plant then shows moving the average temperature away from setpoint, the continuous-control claim fails on its own terms. For the planning claim, a single random FSM within 4-25 states and up to 300 edges on which GPT-4o and GPT-4o-mini, with the paper's exact prompts, fail to find a valid recovery path within five reprompts would overturn the 100% success figure.

Watch

Extended reading notes

Core claim

The paper claims that LLM-based agents become dependable industrial-automation components not because the models are reliable reasoners, but because their proposals pass through a validating environment and a reprompting agent that converts rejection reports into corrective prompts; the authors call this the validator-reprompting loop. Two case studies carry the claim. On 180 randomly generated finite state machines with 4-25 states and up to 300 transitions, action-simulation-validation-reprompting lets GPT-4o and GPT-4o-mini produce a valid recovery path in every instance within five reprompts, with average path deviation under 0.5 steps even on the hardest graphs and mean solve times of a few seconds; the open-source LLaMA-3 (3.2B) baseline fails on many medium and hard instances. On a dual-heater Temperature Control Lab (TCLab) rig under a fan-induced cooling disturbance, the same pattern proposes heater powers on a fixed 30-second cadence and accepts them only when a validator confirms the predicted average temperature moves toward setpoint; GPT-3.5 on the digital twin reaches a time-weighted mean absolute error of 0.62 versus the twin PID's 0.21, the closest of any LLM controller, and a one-off physical comparison lands in the same range as the physical PID. Because the physical platform showed variability and latency that hampered reproducibility, the deterministic comparison runs on a digital twin with fixed thermal parameters (m = 0.004 kg, Cp = 500 J/kg-K, U = 10 W/$m^{2}$-K, emissivity 0.9), and the authors show that removing the reprompting loop degrades control, identifying it as essential for nonlinear dynamics. The paper also catalogs the failure modes that bound the claim: instruction-following lapses, equal-power allocation that ignores state asymmetry, neglected heat losses, and single-step ODE approximations that drift up to 3.3 K from ground truth over 30 seconds.

Load-bearing premise

The load-bearing premise is that the digital twin's fixed thermal parameters (mass 0.004 kg, heat capacity 500 J/kg-K, heat-transfer coefficient 10 W/$m^{2}$-K, emissivity 0.9) represent the physical dual-heater rig accurately enough over each 30-second interval that a control action accepted inside the simulator also moves the real plant toward setpoint.

Editorial extensions

If this is right

  • Rare and unanticipated fault conditions become addressable without retraining: any failure mode expressible as a validator that can reject bad proposals can be fed back into the loop, so coverage grows with the validator set rather than with the model's training data.
  • A single framework can span discrete operating-regime transitions and continuous regulation, so an FSM can serve as the envelope within which an LLM chooses continuous control moves, unifying the two case studies into one control architecture.
  • Closed commercial models (GPT-4o, GPT-4o-mini) currently hold a decisive accuracy-and-latency edge over the open-source LLaMA baseline, meaning near-term deployments will trade data sovereignty and auditability against planning performance.
  • Because the loop is bounded by an iteration budget with a safety override to a fallback policy, LLM autonomy can be contained: unstable behavior is converted into a handoff rather than a runaway.
  • The reprompting loop is required for nonlinear continuous dynamics; ablating it degrades control, so any deployment must keep the validator-reprompter pair rather than relying on the LLM planner alone.

Reading between the lines

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

  • A testable prediction follows from the architecture: as open-source models close the instruction-following gap, the framework's accuracy gap will shrink faster than the raw model gap, because the validator does the correctness work and the LLM only needs to respond to feedback.
  • The fixed 30-second planning interval raises a control-theoretic question the paper does not ask: the loop can only stabilize plants whose time constants exceed roughly the inference-plus-reprompt latency, suggesting an upper bound on process bandwidth for LLM-in-the-loop control that could be quantified.
  • The FSM results measure path length against shortest-path optimality; a natural extension is cost- and risk-weighted transitions (energy, wear, safety-criticality per edge), where the near-optimality the authors observe could translate into economically meaningful recovery policies.
  • The equal-power failure mode suggests that feedback content, not just rejection flags, is the lever: prompts that state 'heater 1 is already the hotter one' should outperform prompts that only report 'temperature criterion not met,' a comparison the paper's own examples make directly testable.
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

4 major / 6 minor

Summary. The paper proposes an agentic control architecture consisting of monitoring, action, simulation, validation, and reprompting agents, and evaluates it in two case studies. In Case Study 1, LLM agents plan recovery paths over randomly generated finite state machines using a deterministic transition checker and up to five reprompting cycles, with the paper reporting 100% valid-path accuracy for GPT-4o and GPT-4o-mini. In Case Study 2, the same framework proposes heater power settings for a dual-heater TCLab system, both on physical hardware and in a Python digital twin, and the results are compared with a PID baseline using TW-MAE and RMSE. The paper concludes that LLM-based agents can unify symbolic planning and continuous control, and it claims that an ablation of the reprompting loop reveals its critical role in handling nonlinear dynamics.

Significance. The FSM part of the paper, if the data inconsistencies are reconciled, is a clean empirical demonstration that a validator–reprompting loop can make LLM symbolic planning reliable on a constrained class of graphs; the deterministic FSM checker and the publicly available code are strengths. The continuous-control contribution is not yet established: the digital twin is never validated against the physical plant, only one model and one physical run are reported, there are no error bars or significance tests, the claimed ablation is absent from the body, and the phrase "similar performance" overstates the measured gaps in Table 5. The manuscript is an interesting architecture study with promising but incomplete evidence for its central claims.

major comments (4)
  1. [Abstract and Section 4.2] The abstract states that "ablation of the reprompting loop reveals its critical role in handling nonlinear dynamics," but no ablation experiment appears in Section 4.2 or elsewhere in the manuscript. Table 5 compares different models and environments but never removes the validation/reprompting loop. This claim should be removed or supported by an actual loop-versus-no-loop comparison.
  2. [Section 4.2.1, Table 5, Appendix B] The digital-twin rows in Table 5 do not demonstrate control competence on the real plant because the twin is never validated. Section 4.2.1 fixes thermal parameters (m=0.004 kg, Cp=500 J/kg-K, U=10 W/m2-K, emissivity 0.9), and Appendix B shows that the Action Agent, Temperature Validator, and reprompters are prompted with the same model and energy-balance equations. Simulation, validation, and "ground truth" therefore use one self-consistent approximation; actions that pass in the twin could transfer poorly to hardware if the 30-second energy-balance prediction is inaccurate. A quantitative comparison of twin predictions with physical TCLab responses is required before the simulation results can support the continuous-control claim.
  3. [Section 4.2.1 and Table 5] The claim of "similar performance" to PID is not supported by the measured point estimates. In the physical system, GPT-3.5 achieves TW-MAE 1.0277 versus 0.7605 for PID, about 35% worse; in the digital twin, GPT-3.5 achieves 0.6182 versus 0.2097 for PID, roughly three times worse. No repeated trials, error bars, or significance tests are reported, and only GPT-3.5 was run on the physical plant. The wording should be revised, or the experiments should be supplemented with replicates and a statistical comparison.
  4. [Section 4.1.2, Tables 1–3, and Abstract] The FSM quantitative evidence contains internal inconsistencies. The abstract says 180 random FSMs, but Section 4.1.2 generates 20 instances for each of the 10 (nodes, edges) rows in Tables 1–3, i.e., 200 FSMs. In addition, under the stated definition of average reprompts as the mean over solved instances, several table values are impossible: Table 2, row (4,4), reports first-pass accuracy 0.75, valid-path accuracy 1.00, and average reprompts 0.00; Table 1, row (4,6), reports first-pass and valid-path accuracy both 0.75 with average reprompts 1.25. The authors should reconcile these numbers and clarify the metric definition.
minor comments (6)
  1. [Table 4] The table reports "Sample Count" but does not define what a sample is; please state whether it is a control interval, an LLM inference call, or an experimental run so that the physical/digital twin counts (53 vs. 148) are interpretable.
  2. [Appendix B] The task prompt for temperature_reprompter_2 says "Adjust the power of heater 1" and "find the heater output for heater 1" although the task is for heater 2, and the agent role for temperature_reprompter_2 is duplicated as "Heater 1."
  3. [Appendix B] The Power Validation Agent prompt says to verify that heaters output "less than or equal to 30" and "greater than or equal to 0"; this should be 0.3 W to match the 0–0.3 W bound used throughout the paper.
  4. [Tables 1–5 and text] Model naming is inconsistent: the paper refers to LLaMA-3 (3.2B), LLaMA-3.2, and LLaMA-3:8b in different places; please unify the notation.
  5. [Section 4.1.2] The "Path quality" metric is defined as deviation from the shortest path, but the paper does not state how the shortest path length is computed for each random FSM; please specify the algorithm (e.g., BFS).
  6. [Section 4.2.1] The sentence about pausing the emulated system during the 30-second planning interval is unclear and contains a typo ("physcial"); please clarify whether the digital twin, the physical plant, or both are paused during LLM inference.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FSM validity checking and digital-twin PID comparisons are external baselines, and no fitted parameter or derivation reduces to its own input.

full rationale

The paper's derivation chain is self-contained. In Case Study 1, valid-path accuracy is determined by the Simulation Agent walking the proposed path against the FSM dictionary, an external ground truth that is independent of the LLM; the reported success rates are not fed back into any fitted parameter. In Case Study 2, the digital twin is an implementation of a stated energy-balance ODE with fixed parameters (m = 0.004 kg, Cp = 500 J/kg-K, U = 10 W/m2-K, emissivity 0.9), and both the PID baseline and the LLM controllers are evaluated against the same simulated plant state, so the TW-MAE/RMSE comparisons are external benchmarks rather than reductions. The temperature validator does accept or reject actions using the same approximate energy-balance model that the Action Agent is prompted to use, but this is internal consistency, not circularity: the reported error metrics are computed from the digital twin's state trajectory, not from the LLM's own predicted temperatures, and no parameter is fitted from LLM outputs. The only self-citations (Vyas and Mercangöz 2024; Gill et al. 2025) position the architecture and prior fault-handling work without supplying a load-bearing theorem or fitted value. The unvalidated fidelity of the digital twin to the physical TCLab is a correctness and generalization risk, and the paper acknowledges broader limitations in Section 5, but neither constitutes a circular derivation.

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

The central results rest on the assumed fidelity of the digital twin and on the representativeness of random FSMs; no constants are fitted to make the claims pass. The missing ablation means the role of the reprompting loop is asserted rather than demonstrated.

assumptions (4)
  • domain assumption Randomly generated connected FSMs are representative of industrial fault-recovery scenarios.
    Used in Section 4.1.2 to benchmark planning; no real plant FSMs or alarm sequences are used, as the discussion section acknowledges.
  • domain assumption The digital twin's fixed energy-balance model (m=0.004 kg, Cp=500 J/kg-K, A=1.2e-3 m^2, U=10 W/m^2-K, emissivity 0.9) accurately represents the physical TCLab heaters over 30-second intervals.
    Section 4.2.1 says a Python digital twin was developed because physical trials were hard to reproduce; control performance in the twin is used to draw conclusions about physical control.
  • domain assumption LLM outputs with temperature=0 and top_p=0.1 are stable enough that repeated runs are comparable.
    Section 3.4 sets sampling parameters 'to favour reproducibility over creativity', but no repeated-run variance is reported.
  • ad hoc to paper The validator's monotonic acceptance rule (new action must move predicted average temperature closer to target) leads to closed-loop convergence.
    Implicit in the temperature validation agent described in Appendix B; no proof of convergence is offered, and the rule is judged on the LLM's own predicted temperatures.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Autonomous Control Leveraging LLMs: An Agentic Framework for Next-Generation Industrial Automation." pith.science (2026). https://pith.science/paper/MIE4N62L

@misc{pith2026250707115,
  author       = {Pith},
  title        = {Pith review of: Autonomous Control Leveraging LLMs: An Agentic Framework for Next-Generation Industrial Automation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MIE4N62L}},
  note         = {Machine review of arXiv:2507.07115}
}
read the original abstract

The increasing complexity of modern chemical processes, coupled with workforce shortages and intricate fault scenarios, demands novel automation paradigms that blend symbolic reasoning with adaptive control. In this work, we introduce a unified agentic framework that leverages large language models (LLMs) for both discrete fault-recovery planning and continuous process control within a single architecture. We adopt Finite State Machines (FSMs) as interpretable operating envelopes: an LLM-driven planning agent proposes recovery sequences through the FSM, a Simulation Agent executes and checks each transition, and a Validator-Reprompting loop iteratively refines invalid plans. In Case Study 1, across 180 randomly generated FSMs of varying sizes (4-25 states, 4-300 transitions), GPT-4o and GPT-4o-mini achieve 100% valid-path success within five reprompts-outperforming open-source LLMs in both accuracy and latency. In Case Study 2, the same framework modulates dual-heater inputs on a laboratory TCLab platform (and its digital twin) to maintain a target average temperature under persistent asymmetric disturbances. Compared to classical PID control, our LLM-based controller attains similar performance, while ablation of the prompting loop reveals its critical role in handling nonlinear dynamics. We analyze key failure modes-such as instruction following lapses and coarse ODE approximations. Our results demonstrate that, with structured feedback and modular agents, LLMs can unify high-level symbolic planningand low-level continuous control, paving the way towards resilient, language-driven automation in chemical engineering.

Figures

Figures reproduced from arXiv: 2507.07115 by the authors.

Figure 1
Figure 1. Five-component agentic pipeline: monitoring, action, simulation, validation, and reprompting. 5 [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Illustration of a Finite State Machine encoded as a dictionary [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Agent setup for FSM case study. The framework consists of the following components: • Action Agent: Given the initial and goal states, this agent proposes a candidate path composed of intermediate FSM states. The agent operates purely via natural language and symbolic reasoning, without access to ground-truth search algorithms or environment dynamics. • Simulation Agent: Serves as a lightweight path executor by walk… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Path accuracy comparison across models and FSM sizes. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Multi-agent framework for continuous temperature control. [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Physical experimental setup for dual-heater temperature control. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: Three configurations used for performance evaluation. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: PID controller vs GPT 3.5 - Physical System [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Control performance of the system - Simulated in digital environment [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. ADMITBench: A Safety-Governed Reference Framework for Evaluating the Admissibility of Industrial LLM Advisories

    cs.AI 2026-08 conditional novelty 6.0 of 10

    ADMITBench defines and implements a safety-governed evaluation framework that judges LLM industrial advisories at the level of the proposed action, using non-compensatory eligibility gates and versioned plant profiles.

  2. Safety-Gated Agentic Supervisory Control on a Coupled Distillation Benchmark: Regime Map, Auditable Gate, and Co-Design Findings

    eess.SY 2026-07 conditional novelty 6.0 of 10

    On a simulated distillation column, an ungated LLM supervisor beats linear MPC at off-nominal target re-planning but is ~10–16x worse at disturbance rejection; a forked-twin safety gate contains failures, yet 534 of 5...

  3. Towards an Intention Abstraction Layer for Autonomous Industrial Systems

    cs.AI 2026-07 conditional novelty 5.0 of 10

    A runtime middleware records autonomous agents' goals as structured, ontology-linked intentions and flags conflicts between them before execution, demonstrated on a two-agent production/energy scenario.

  4. Closed-Loop Control with Rule-Aligned Small Language Models and Multi-Agent Self-Correction

    cs.AI 2026-06 conditional novelty 5.0 of 10

    A GRPO-aligned Qwen2.5-1.5B multi-agent loop reaches 91.5% expert-action agreement and 95% in-range thermal regulation in randomized first-order simulations at 3.84 s mean latency.

Reference graph

Works this paper leans on

33 extracted references · 24 canonical work pages · cited by 4 Pith papers

  1. [1]

    Agrawal, P., Vasania, S., and Tan, C. (2024). Can LLMs perform structured graph reasoning? arXiv:2402.01805 [cs]

    work page Pith review arXiv 2024

  2. [2]

    T., Gopaluni, B., Mercangöz, M., Pantelides, C

    Baldea, M., Georgiou, A. T., Gopaluni, B., Mercangöz, M., Pantelides, C. C., Sheth, K., Zavala, V. M., and Georgakis, C. (2025). From automated to autonomous process operations. Computers & Chemical Engineering , 196:109064

    work page 2025

  3. [3]

    Besta, M., Blach, N., Kubicek, A., Gerstenberger, R., Podstawski, M., Gianinazzi, L., Gajda, J., Lehmann, T., Niewiadomski, H., Nyczyk, P., and Hoefler, T. (2024). Graph of thoughts: Solving elaborate problems with large language models. Proceedings of the AAAI Conference on Artificial Intelligence , 38(16):17682–17690

    work page 2024

  4. [4]

    Borghesan, F., Zagorowska, M., and Mercang \"o z, M. (2022). Unmanned and autonomous systems: Future of automation in process and energy industries. IFAC-PapersOnLine , 55(7):875--882

    work page 2022

  5. [5]

    Buddaraju, D. (2011). Performance of control room operators in alarm management . Louisiana State University and Agricultural & Mechanical College

    work page 2011

  6. [6]

    Chase, H. (2022). LangChain

    2022

  7. [7]

    Chen, Q., Garayev, M., and Liu, D. (2024). State-based fault diagnosis of finite-state vector discrete-event systems via integer linear programming. Sensors , 25(5):1452

    work page 2024

  8. [8]

    Crewai: An autonomous control framework

    CrewAI (2024). Crewai: An autonomous control framework. Accessed: 2024-11-08

    work page 2024

Show all 33 references
  1. [9]

    Q., Ogundare, O., and Faruque, M

    Fakih, M., Dharmaji, R., Moghaddas, Y., Araya, G. Q., Ogundare, O., and Faruque, M. A. (2024). Llm4plc: Harnessing large language models for verifiable programming of plcs in industrial control systems. In Proc. 46th International Conference on Software Engineering: Software E...

    2024

  2. [10]

    S., Vyas, J., Markaj, A., Gehlhoff, F., and Mercang "o z, M

    Gill, M. S., Vyas, J., Markaj, A., Gehlhoff, F., and Mercang "o z, M. (2025). Leveraging llm agents and digital twins for fault handling in process plants. arXiv preprint arXiv:2505.02076

    2025

  3. [11]

    Goel, P., Datta, A., and Mannan, M. S. (2017). Industrial alarm systems: Challenges and opportunities. Journal of Loss Prevention in the Process Industries , 50:23--36

    2017

  4. [12]

    Hou, Z. (2025). Neural-symbolic reasoning: Towards the integration of logical reasoning with large language models. Authorea Preprints

    2025

  5. [13]

    Jin, B., Liu, G., Han, C., Jiang, M., Ji, H., and Han, J. (2024). Large Language Models on Graphs : A Comprehensive Survey . IEEE Transactions on Knowledge and Data Engineering , 36(12):8622--8642

    2024

  6. [14]

    2.1 million manufacturing jobs could go unfilled by 2030

    National Association of Manufacturers (2021). 2.1 million manufacturing jobs could go unfilled by 2030. Accessed: 2025-05-26

    2021

  7. [15]

    Oliveira, P. M. and Hedengren, J. D. (2019). An apmonitor temperature lab pid control experiment for undergraduate students. In 2019 24th IEEE International Conference on Emerging Technologies and Factory Automation (ETFA) , pages 790--797. IEEE

    2019

  8. [16]

    Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., et al. (2022). Training language models to follow instructions with human feedback. Advances in neural information processing systems , 35:27730--27744

    2022

  9. [17]

    Provost, J., Roussel, J.-M., and Faure, J.-M. (2011). Testing programmable logic controllers from finite state machines specification. In 2011 3rd International Workshop on Dependable Control of Discrete Systems , pages 1--6

    2011

  10. [18]

    Sheng, X., Yu, C., Cui, X., and Zhou, Y. (2024). Large language model and digital twins empowered asynchronous federated learning for secure data sharing in intelligent labeling. Mathematics , 12(22):3550

    2024

  11. [19]

    Shinn, N., Cassano, F., Berman, E., Gopinath, A., Narasimhan, K., and Yao, S. (2023). Reflexion: Language agents with verbal reinforcement learning

    2023

  12. [20]

    H., Wu, J., Washington, C., Sadler, B

    Song, C. H., Wu, J., Washington, C., Sadler, B. M., Chao, W.-L., and Su, Y. (2023). Llm-planner: Few-shot grounded planning for embodied agents with large language models. In Proceedings of the IEEE/CVF international conference on computer vision , pages 2998--3009

    2023

  13. [21]

    Creating pathways for tomorrow’s workforce today: Beyond reskilling in manufacturing

    The Manufacturing Institute and Deloitte (2021). Creating pathways for tomorrow’s workforce today: Beyond reskilling in manufacturing. Accessed: 2025-05-26

    2021

  14. [22]

    Valmeekam, K., Olmo, A., Sreedharan, S., and Kambhampati, S. (2022). Large language models still can't plan (a benchmark for llms on planning and reasoning about change). In NeurIPS 2022 Foundation Models for Decision Making Workshop

    2022

  15. [23]

    and Mercang\" o z, M

    Vyas, J. and Mercang\" o z, M. (2024). Autonomous industrial control using an agentic framework with large language models

    2024

  16. [24]

    Wan, G., Wu, Y., Chen, J., and Li, S. (2024). Reasoning aware self-consistency: Leveraging reasoning paths for efficient llm sampling. arXiv preprint arXiv:2408.17017

    2024 arXiv

  17. [25]

    V., and Zhou, D

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., ichter, b., Xia, F., Chi, E., Le, Q. V., and Zhou, D. (2022). Chain-of-thought prompting elicits reasoning in large language models. In Koyejo, S., Mohamed, S., Agarwal, A., Belgrave, D., Cho, K., and Oh, A., editors, Advances in N...

    2022

  18. [26]

    Wu, Q., Bansal, G., Zhang, J., Wu, Y., Li, B., Zhu, E., Jiang, L., Zhang, X., Zhang, S., Liu, J., et al. (2023). Autogen: Enabling next-gen llm applications via multi-agent conversation. arXiv preprint arXiv:2308.08155

    2023 arXiv

  19. [27]

    Xia, Y., Jazdi, N., Zhang, J., Shah, C., and Weyrich, M. (2024). Control industrial automation system with large language models. arXiv preprint arXiv:2409.18009

    2024 arXiv

  20. [28]

    Xia, Y., Shenoy, M., Jazdi, N., and Weyrich, M. (2023). Towards autonomous system: flexible modular production system enhanced with large language model agents. In Proc. 2023 IEEE 28th International Conference on Emerging Technologies and Factory Automation (ETFA) , pages 1--8

    2023

  21. [29]

    Xu, W., Banburski, A., and Jojic, N. (2024). Reprompting: Automated chain-of-thought prompt inference through gibbs sampling

    2024

  22. [30]

    Yang, L., Luo, S., Cheng, X., and Yu, L. (2025). Leveraging large language models for enhanced digital twin modeling: Trends, methods, and challenges. arXiv preprint arXiv:2503.02167

    2025 arXiv

  23. [31]

    L., Cao, Y., and Narasimhan, K

    Yao, S., Yu, D., Zhao, J., Shafran, I., Griffiths, T. L., Cao, Y., and Narasimhan, K. (2023). Tree of thoughts: Deliberate problem solving with large language models

    2023

  24. [32]

    Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., and Cao, Y. (2022). React: Synergizing reasoning and acting in language models

    2022

  25. [33]

    Zhang, X., Zhao, Y., Lin, H., Riaz, S., and Elahi, H. (2021). Real-time fault diagnosis and fault-tolerant control strategy for hall sensors in permanent magnet brushless dc motor drives. Electronics , 10(11):1268

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.