Multi-Agent LLM Architectures for Quality Assurance and Governance in Personalized Learning Systems

Thesis (MA)

Advisor(s): Alisa Mehler(alisa.mehler(at)tum.de)

CONTEXT

As personalized learning systems become increasingly automated, new risks emerge: hallucinated content, inconsistencies, didactic errors, and loss of governance control. Ensuring trustworthiness and robustness becomes critical when LLMs generate educational material dynamically.

One promising approach is the use of multi-agent LLM architectures, where specialized agents collaborate in structured workflows such as Generate–Verify–Refine.

This thesis investigates how multi-agent systems can ensure content quality, didactic correctness, and governance compliance in personalized learning environments based on the textbook Informationsmanagement by Krcmar (2015).

The objective is to design and implement a governance-oriented architecture that reduces hallucination risks, increases transparency, and enables scalable quality assurance mechanisms suitable for academic and executive education.

Research Objectives of the master’s thesis could include:

• Definition of specialized agent roles (generation, review, QA, governance)
• Multi-stage validation workflows (Generate–Verify–Refine)
• Prompt governance mechanisms and quality criteria
• Comparison of single-agent vs. multi-agent architectures
• Human-in-the-loop integration for critical learning content

 Potential Research Questions guiding the thesis are:

This topic can be grounded in research questions such as:

• What factors determine the trustworthiness of automatically generated learning content?
• How can governance mechanisms ensure quality and consistency in adaptive AI-based education?
• What barriers hinder scalable validation of LLM-generated educational material?
• How can support and verification structures be strengthened through multi-agent architectures?
• Which workflows best reduce hallucination risks and increase transparency?

These and similar questions can serve as a foundation for a BA/MA/Guided Research thesis.

Other research questions to be pursued in the thesis can be suggested and discussed.

 

TASK(S)

Potential thesis activities include:

• Review research on multi-agent LLM systems and AI governance
• Identify risks of automated educational content generation
• Define agent roles for content creation, didactic review, and validation
• Design multi-stage agent workflows for robust learning path generation
• Implement a prototype comparing single-agent vs. multi-agent approaches
• Develop evaluation criteria for trustworthiness and quality assurance
• Discuss governance frameworks for scalable AI-based education

• Explore human oversight strategies and maintainability implications

   

REQUIREMENTS

• Strong interest in AI architectures, governance, and educational systems
• Experience with Python and LLM-based prototyping
• Motivation to explore robustness, QA workflows, and system control
• Structured approach to evaluating trust and quality in AI systems
• Independent and reliable work style