Large Language Models as Managing Systems
|
Typ | Masterarbeit |
|---|---|---|
| Aushang | LLM4SAS M Sc Thesis Proposal.pdf | |
| Betreuer | Wenden Sie sich bei Interesse oder Fragen bitte an: Vincenzo Scotti (E-Mail: vincenzo.scotti@kit.edu) |
Motivation
This thesis is motivated by the potential to enhance Self-Adaptive Systems (SAS), which manage themselves through monitoring, analysis, planning, and execution. Traditionally, these systems rely on engineered components within the MAPE loop. However, recent advancements in Large Language Models (LLMs) such as ChatGPT, DeepSeek, and Gemini suggest that these models may offer strong capabilities in reasoning and decision-making. This work explores whether LLMs can serve as core components or even end-to-end controllers within adaptive systems. The goal is to evaluate their performance, flexibility, and limitations in real-world scenarios.
Tasks
The goal of this thesis is to investigate the viability of using LLMs as managing systems in the context of SAS. The student will explore two main approaches: (i) end-to-end usage of LLMs for predicting management actions, and (ii) a modular approach where the LLM serves one or more specific functions. The research will involve designing experiments to test LLMs under varying conditions, evaluating performance, and assessing the benefits of fine-tuning. A comparative analysis against traditional management components will help determine where LLMs add value and where limitations remain.
Tools/Technology
Python, HuggingFace Transformers, Llama CPP, PEFT
Benefits
- Work with advanced technologies: Gain experience with LLMs and adaptive systems.
- Research integration: Aligns with ongoing work in AI-based system management.
- Close supervision: Regular guidance and access to relevant expertise