How to Apply

Next Steps

• If Accepted:

  • We will formalize your topic and discuss project goals.
  • Ensure you meet university-specific requirements (e.g., registration, defense talks).

• If Declined:

  • You may revise your topic or explore alternative supervisors.

Evaluation Criteria

Your thesis will be graded on:

• Literature review.
• Scientific approach and methodology.
• Clear structure and comprehensive documentation.
• Novelty and significance (especially?for master students).
• Quality of implementation or study design.

Feel free to contact me for further clarification or to apply. Looking forward to working on exciting projects together!

?

?

Open Topics:

Training Reliable In-Car Assistant Agents

Current LLM agents prioritize task completion over compliance: they fabricate capabilities, act on ambiguous requests without clarifying, and violate domain policies. In our CAR-bench benchmark [1], even frontier models achieve less than 54% consistent success across these dimensions. This thesis targets one or more of these failure modes by training small language models (SLMs) within the CAR-bench environment, using paradigms such as self-evolving agents, LoRA fine-tuning, RLVR, or supervised fine-tuning. Focus area and method can be tailored to the student's interests. Familiarity with LLM fine-tuning or reinforcement learning is beneficial but not required. I am also open to self-proposed topics within the CAR-bench environment.

[1]?Kirmayr, Johannes, Lukas Stappen, and Elisabeth André. "CAR-bench: Evaluating the Consistency and Limit-Awareness of LLM Agents under Real-World Uncertainty."?arXiv preprint arXiv:2601.22027?(2026).

?

Synthetic Multi-Turn Data Generation for Training Tool-Using In-Car Voice Agents

Training reliable tool-using agents requires large-scale interaction data that is expensive to collect manually. This thesis explores novel synthetic data generation approaches for the CAR-bench environment [1], building on work like APIGen-MT [2] which used synthetic pipelines to train small models that outperform frontier LLMs on τ-bench. The student will research and develop methods to generate diverse, verifiable multi-turn trajectories across CAR-bench's 58 tools and three task types (base, hallucination, disambiguation), with a focus on data quality, verification, and downstream training effectiveness for SLM agents. Familiarity with LLM, graph networks, and data pipelines is beneficial but not required.

?

[1]?Kirmayr, Johannes, Lukas Stappen, and Elisabeth André. "CAR-bench: Evaluating the Consistency and Limit-Awareness of LLM Agents under Real-World Uncertainty."?arXiv preprint arXiv:2601.22027?(2026).

[2]?Prabhakar, Akshara, et al. "Apigen-mt: Agentic pipeline for multi-turn data generation via simulated agent-human interplay."?arXiv preprint arXiv:2504.03601?(2025).

?

Human-Agent Interaction for Multi-Step LLM Assistants

As LLM agents handle increasingly complex, multi-step tasks, new HCI challenges arise: How do users maintain trust when agents execute long-running operations? How should agents communicate uncertainty, progress, or errors? This thesis investigates human-agent interaction design for multi-turn, tool-using voice assistants in the automotive domain. Possible research directions include user trust dynamics during multi-step task execution, interaction design patterns for agent transparency and error recovery, or the effect of agent communication strategies on user experience. The thesis will likely include a user study. Focus area can be tailored to the student's interests. Background in LLMs, HCI, UX research, or study design is beneficial but not required.