Department of Mathematics and Computer Science

Large Language Models enable conversational agents to adjust their communication style to individual users. In sensitive contexts such as healthcare, this creates opportunities for AI systems to communicate in ways that better match users' individual preferences. Previous work in our research group has explored a patient-facing AI companion for supporting patients while waiting in an emergency department.

However, customization itself introduces a design challenge. Users may differ not only in how they want an AI to communicate, but also in how much effort and control they want to invest in configuring it. Some users may prefer directly specifying their preferences, while others may prefer suggestions, lightweight adjustments during an ongoing conversation, or greater system initiative. A single customization mechanism may therefore not accommodate all users equally well.

This thesis explores flexible customization: interaction techniques that allow users to access different levels of control while keeping the effort required for customization low. Rather than treating customization as a separate configuration task before using an AI system, the thesis investigates how customization can become part of the normal interaction with a conversational agent.

Possible interaction concepts include lightweight style tags, customization based on feedback to individual responses, switching between coarse and fine-grained controls, or previews that immediately show how changing a preference would alter an AI-generated response. 

Procedure

The thesis builds on an existing LLM-based conversational agent and findings from an ongoing research project on customizable AI companions for the emergency department.

First, relevant literature on customization, mixed-initiative interaction, controllability, and Human-AI interaction is reviewed. Existing project findings and prototypes are then analyzed to derive requirements for flexible customization.

Based on these requirements, several interaction concepts are explored and iteratively implemented as an interactive LLM-based prototype. The prototype should allow users to customize the conversational agent at different levels of control without requiring them to formulate complex prompts.

Finally, the prototype is evaluated in a user study. The evaluation investigates how participants use the available levels of customization and examines outcomes such as perceived control, mental effort, interaction time, usability, and perceived fit of the resulting communication style. Findings are used to derive design recommendations for flexible customization mechanisms in future LLM-based conversational agents.  

References

• Klein, J., Sörries, P., Mutlugil, Y., George, C., Altendorf, M., Möckel, M., & Müller-Birn, C. (2026). With, not For: Co-Designing a Patient-Facing AI Companion Concept for the Emergency Department Waiting Area. Proceedings of CHI '26. 
• Zhang, R., Ringland, K. E., Paan, M., Mohr, D. C., & Reddy, M. (2021). Designing for Emotional Well-being: Integrating Persuasion and Customization into Mental Health Technologies. Proceedings of CHI '21.
• Gmeiner, F., et al. (2025). Intent Tagging: Exploring Micro-Prompting Interactions for Supporting Granular Human-GenAI Co-Creation Workflows. Proceedings of CHI '25.
• Riche, N., et al. (2025). AI-Instruments: Embodying Prompts as Instruments to Abstract & Reflect Graphical Interface Commands as General-Purpose Tools. Proceedings of CHI '25.
• Vossen, W., Szymanski, M., & Verbert, K. (2024). The Effect of Personalizing a Psychotherapy Conversational Agent on Therapeutic Bond and Usage Intentions. Proceedings of IUI '24.