Large Language Models for Self-Adaptive Systems: Feedback-Loop Integration and Agentic Adaptation

Self-Adaptive Systems (SAS) address uncertainty and change through feedback loops that monitor the system and its environment, analyze possible adaptation options, plan suitable adaptations, and execute them to achieve system goals.
The pervasive uprising of Large Language Model (LLM)-based technologies introduces new opportunities to enhance decision-making in SAS.
However, so far, the integration of LLMs raises fundamental architectural questions that have not been sufficiently explored.
In this paper, we investigate two architectural paradigms for LLM-driven self-adaptation to gain a practical understanding about whether LLMs should be embedded into existing feedback-loop structures or employed as autonomous, end-to-end adaptive agents.
The first paradigm integrates an LLM into the classical MAPE-K feedback loop by replacing the analysis and planning components while preserving explicit monitoring, execution, and knowledge models.
The second paradigm adopts an agentic approach in which an LLM operates as an autonomous decision-maker, implicitly realizing the adaptation loop in an end-to-end manner.
We instantiate both paradigms on two SAS exemplars to empirically compare them with respect to adaptation quality, stability, responsiveness, and transparency. ... mehr
Our results highlight clear trade-offs between the two approaches.
While the LLM-enhanced MAPE-K architecture offers greater controllability and explainability, the agentic approach provides increased flexibility at the cost of reduced architectural clarity and predictability.