[論文レビュー] When control meets large language models: From words to dynamics

Komeil Nosrati, Aleksei Tepljakov+2 more|arXiv (Cornell University)|Feb 3, 2026

Topic Modeling被引用数 0

ひとこと要約

本文は英語のままの長文要約は保持します。

ABSTRACT

While large language models (LLMs) are transforming engineering and technology through enhanced control capabilities and decision support, they are simultaneously evolving into complex dynamical systems whose behavior must be regulated. This duality highlights a reciprocal connection in which prompts support control system design while control theory helps shape prompts to achieve specific goals efficiently. In this study, we frame this emerging interconnection of LLM and control as a bidirectional continuum, from prompt design to system dynamics. First, we investigate how LLMs can advance the field of control in two distinct capacities: directly, by assisting in the design and synthesis of controllers, and indirectly, by augmenting research workflows. Second, we examine how control concepts help LLMs steer their trajectories away from undesired meanings, improving reachability and alignment via input optimization, parameter editing, and activation-level interventions. Third, we look into deeper integrations by treating LLMs as dynamic systems within a state-space framework, where their internal representations are closely linked to external control loops. Finally, we identify key challenges and outline future research directions to understand LLM behavior and develop interpretable and controllable LLMs that are as trustworthy and robust as their electromechanical counterparts, thereby ensuring they continue to support and safeguard society.

研究の動機と目的

• Trace the historical evolution of AI–control synergy from cybernetics to modern LLMs.
• Categorize interactions between LLMs and control into three paradigms: indirect support, direct design guidance, and internal dynamical analysis.
• Analyze how control concepts can steer LLMs toward reliable, aligned outputs.
• Propose the use of structured state-space models (SSMs) to study LLM internal dynamics and controllability.
• Outline challenges and future research directions toward trustworthy, robust controllable LLMs.

提案手法

• Frame a bidirectional continuum between prompts, control, and system dynamics.
• Classify LLM–control interactions into three paradigms: (a) LLMs augment control research and directly assist design, (b) control concepts shape LLM outputs via input optimization, editing, and interventions, (c) LLMs treated as dynamical systems analyzed with SSMs.
• Discuss historical evolution from cybernetics and reinforcement learning to RLHF and modern LLM applications.
• Propose a levels-based framework (L0–L2) for LLM-augmented control workflows and provide example listings.
• Describe LLM-based tooling for control design, including feedback tuning, zero-shot and few-shot prompting, CoT prompting, and agentic architectures.
• Highlight the potential of LLMs to assist in controller tuning (offline and online) and in optimization within control pipelines.

Figure 1: Capabilities of LLMS.

実験結果

リサーチクエスチョン

• RQ1 Why is the intersection of LLMs and control important?
• RQ2 When did the connection between LLMs and control emerge and how has it evolved?
• RQ3 Where do control concepts support LLMs and where do LLMs support control?
• RQ4 What are the challenges and future trends at the LLM–control intersection?

主な発見

• We identify three main interaction paradigms between LLMs and control systems: indirect workflow augmentation, direct control-design assistance, and internal dynamical analysis using state-space concepts.
• LLMs can assist control research and workflow tasks by automating data processing, literature synthesis, and experiment design, structured into three capability levels (L0–L2).
• LLMs can directly participate in control design, including feedback tuning, optimization, symbolic reasoning, design-space exploration, and data-driven modeling, using zero-shot, few-shot, and Chain-of-Thought prompting.
• Control concepts can shape LLM behavior to improve reliability and alignment via input optimization, model editing, and activation-level interventions.
• LLMs can be modeled as dynamic systems (state-space) to study controllability, observability, and stability, linking internal representations to external control loops.
• Historically, the AI–control synergy traces from cybernetics and RL to modern RLHF-enabled LLMs, RL, and Transformer-based architectures, with recent work integrating state-space ideas (e.g., Mamba) and agentic frameworks (e.g., ControlAgent, SmartControl).

Figure 2: Conceptual map of three key intersections between control and LLMs.