AI Control Scientist: LLM-driven Agentic System for Automated Control Design

Haiteng Wang; Weihao Li; Jing Zhang; Lei Ren

cs.AIAug 27, 2026

AI Control Scientist: LLM-driven Agentic System for Automated Control Design

Authors: Haiteng Wang, Weihao Li, Jing Zhang, Lei Ren

Abstract

Control system design is critical for modern industry, such as chemical process temperature regulation and aero-engine control. However,traditional control design workflows rely heavily on expert knowledge and extensive manual parameter tuning, resulting in limited efficiency and scalability. To this end, this paper proposes AI Control Scientist (AICS), the first large language model (LLM)-driven agent capable of automatically generating optimized controller from language design requirements. Specifically, a Task Modeling Agent interprets user requirements to engineering constraints; a Controller Design Agent generate candidate controller structures and executable code; and a Parameter Tuning Agent refine controller parameters under closed-loop performance criteria. Experiments demonstrate that the proposed agentic system can automatically generate multiple representative control systems, outperforms existing automated baselines in both design success rate and optimization efficiency. This work has the potential to transform control system design from human-driven to agent-driven, paving the way for model predictive control and other advanced control systems design.

Figures & tables

Fig. 1: End-to-end execution example of the proposed LLM-driven agentic system for an industrial boiler control task. The system automatically transforms natural-language engineering requirements into rigorous mathematical constraints and outputs an optimized control strategy validated through closed-loop simulation.

Fig. 2: The overall architecture of the proposed Multi-Agent System. It consists of a Task Modeling Agent for semantic formalization that translates ambiguous natural-language requirements into rigorous mathematical constraints, a Controller Design Agent for for logic-grounded code generation that synthesizes theoretically sound controller structures based on plant characteristics, and a Parameter Tuning Agent for precision optimization that leverages external numerical solvers to refine control gains under closed-loop stability criteria.

Methods	First-ord Stable		Second-ord Stable		First-ord Unstable		Second-ord Unstable		With Delay		Higher Order
Methods	Rate	Iter.	Rate	Iter.	Rate	Iter.	Rate	Iter.	Rate	Iter.	Rate	Iter.
AICS	98%	1.20	98%	1.45	100%	1.00	90%	2.44	100%	1.00	74%	2.32
ControlAgent	98%	1.96	98%	2.03	98%	3.17	74%	6.83	96%	1.06	54%	2.00
Zero-shot	30%	/	0%	/	22%	/	0%	/	62%	/	8%	/
Zero-shot with feedback	70%	2.69	4%	6.50	50%	2.92	2%	5.00	100%	1.78	20%	3.70
Few-shot	30%	/	6%	/	10%	/	0%	/	18%	/	4%	/

TABLE I: Comparison of passing rates (%) among different control design methods across various dynamic systems. Our proposed agentic system consistently outperforms baseline methods, particularly in complex unstable and higher-order systems.

Fig. 3: Time-response evolution during the parameter tuning process. The proposed system (AICS, top row) rapidly explores the parameter space and converges to perfect reference tracking, whereas the baseline (ControlAgent, bottom row) gets trapped in a local optimum with severe steady-state error.

Fig. 4: Boxplot of iteration counts for successful trials. Ours demonstrates significantly lower medians, indicating faster and more stable convergence.

Model Backbone	Params	First-ord Stable		Second-ord Stable		First-ord Unstable		Second-ord Unstable		With Delay		Higher Order
Model Backbone	Params	Rate	Iter.	Rate	Iter.	Rate	Iter.	Rate	Iter.	Rate	Iter.	Rate	Iter.
Qwen3-8B	8B	98%	1.39	98%	2.80	98%	1.43	74%	2.24	96%	1.12	54%	2.07
GPT-4o-mini	$\sim$ 8B	90%	1.07	96%	1.92	96%	1.00	82%	2.56	96%	1.02	68%	3.44
Qwen3-32B	32B	96%	1.19	100%	3.20	96%	1.04	84%	2.31	92%	1.04	66%	2.48
GLM-4.7-Flash	30B	94%	1.17	92%	1.65	88%	1.09	86%	1.72	92%	1.07	70%	2.23
Qwen3-235B-A22B	235B	100%	1.66	98%	2.06	100%	1.66	98%	1.98	98%	1.33	74%	3.38

TABLE II: Performance comparison of different LLM backbones across various dynamic systems (RQ3). The table records the final pass rate (%) within 10 iterations and the average number of iterations (Iter.) for successful samples.

Fig. 5: Ablation passing rate gap across different system types. The area collapse of the green line illustrates the critical role of numerical optimization in complex systems.

Fig. 6: A comprehensive case study of the proposed system. It illustrates the transparent and interpretable workflow across Semantic Formalization (Stage 1), Code Generation with error self-correction (Stage 2), and Simulation-driven Parameter Tuning (Stage 3).

Fig. 7: Passing rate versus iteration budget. The proposed method (Ours) achieves faster convergence and a higher performance upper bound compared to the ablated variants.

AI Control

Artificial intelligence control studies how to direct AI agents and use AI to control robots and other physical systems reliably.

See all 66 papers

An LLM-Driven Workflow for Automated Process Control Strategy Generation and Tuning from Dynamic Process Models
Jul 23, 2026Ari Luna Rueda, Eike Cramer, Klaus Hellgardt +1Language Model-Based Control Feedback Control
Department of Chemical Engineering Imperial College London London, UK · Department of Chemical Engineering University College London London, UK
AgentFactory: Towards Automated Agentic System Design and Optimization
Sep 1, 2026Enci Zhang, Haofeng Wang, Yuesheng Zhu +2Large Language Model-Guided Optimization Agentic Workflows
Peking University Shenzhen Graduate School, Shenzhen, China
Supervising Ralph Wiggum: Exploring a Metacognitive Co-Regulation Agentic AI Loop for Engineering Design
Mar 25, 2026Zeda Xu, Nikolas Martelaro, Christopher McCombLLM Agent Workflow Optimization Metacognition in Language Models
Department of Mechanical Engineering, Carnegie Mellon University, Pittsburgh, PA 15213 · Human-Computer Interaction Institute, Carnegie Mellon University, Pittsburgh, PA, 15213 · Department of Mechanical Engineering, Carnegie Mellon University, Pittsburgh, PA, 15213

Jul 23, 2026J/K move · Enter open · S save

An LLM-Driven Workflow for Automated Process Control Strategy Generation and Tuning from Dynamic Process Models

Ari Luna Rueda, Eike Cramer, Klaus Hellgardt, Mehmet Mercangöz

Department of Chemical Engineering Imperial College London London, UK · Department of Chemical Engineering University College London London, UK

Language Model-Based Control Feedback Control LLM Code Generation

Open paper AskPDF arXiv

We present a structured large-language-model-driven workflow for automated multi-variable control design from dynamic process models. The workflow decomposes the design task into constrained code-generation steps: plant-interface construction, normalization, manipulated-variable controlled-variable (MV-CV) pairing, controller specification, closed loop simulation, scenario generation, performance evaluation and Bayesian-optimization (BO) based tuning. Generated artifacts are executed and validated before downstream tasks proceed, and failed artifacts are repaired using validation feedback. The proposed approach is demonstrated on a nonlinear gas-preheater benchmark with coupled pressure and temperature dynamics. The generated workflow produces a physically consistent decentralized PI (proportional-integral) feedback-feedforward control structure and an executable tuning environment. Bayesian optimization reduces the closed loop performance objective, which aggregates set-point tracking and disturbance-rejection errors for the controlled variables, by approximately 26.5% relative to the initial controller generated by the workflow, mainly through improved pressure-loop transient performance. This figure quantifies the automated tuning stage rather than a comparison against a manually designed controller. The results demonstrate the feasibility of using structured large-language-model-based code generation to construct executable control-design workflows, while also highlighting the need for broader validation on larger plantwide-control benchmarks.

Abstract

AI Control

Explore similar work

An LLM-Driven Workflow for Automated Process Control Strategy Generation and Tuning from Dynamic Process Models

AgentFactory: Towards Automated Agentic System Design and Optimization

Supervising Ralph Wiggum: Exploring a Metacognitive Co-Regulation Agentic AI Loop for Engineering Design