Max Mittelstädt
Ece Sutanrikulu
Lumbardh Ljatifi
Julia Geltl
Berfin Berg
Ann-Marie Atzkern
Anusha Kanagarasa
Vanchha Chandrayan
Ignacio Alvarez
As part of a university project at the Technische Hochschule Ingolstadt, we developed and evaluated an in-vehicle conversational agent that uses Large Language Models (LLMs) to detect driver stress and generate personalized emotional support. The goal was to investigate whether incorporating a Driver Personality Profile (DPP) into the system's prompts improves perceived support, emotional presence, and user acceptance. The results and methodology were compiled into a research paper, which has been officially accepted at the AutoUI 26 conference.
Research, Ideation, User research, Prototyping, Data analysis
We asked: To what extent does providing an LLM agent with individual personality dimensions (through a Driver Personality Profile (DPP)) influence the perceived relevance and acceptance of its emotional support responses? To find out, we compared a baseline voice assistant against a DPP-informed variant across two empirical studies. Surprisingly, acceptance ratings showed no detectable difference, and drivers couldn't tell them apart. This points to output anchoring: strict brevity and safety constraints overpowered the personality prompt, anchoring both models to the same generic register.
Driver states such as stress, anxiety, and fatigue heavily compromise driving performance and safety. While modern Driver Monitoring Systems successfully detect these emotional cues, they usually focus on rigid warnings rather than adaptive, individually tailored feedback. This project explored how Large Language Models (LLMs) can bridge this gap by delivering context-aware, empathetic responses designed to maintain the driver's emotional balance.
To test this, we developed an in-vehicle simulation pipeline using the CARLA simulator and ScenarioRunner to induce controlled stress through traffic congestion and time pressure. The system integrated Whisper for speech transcription, Ollama (llama2:7b-chat) for text generation, and XTTS for speech rendering. We implemented a comparative design evaluating a non-personalized baseline agent against a Driver Personality Profile (DPP) condition, where psychological traits derived from standardized questionnaires were injected into the system prompts to dynamically adapt the agent's tone.
We assessed the system through a laboratory simulator study (n=14) and an online video pilot (n=12). Surprisingly, quantitative metrics revealed no statistically significant differences in perceived usefulness or social-emotional presence between the conditions. Through lexical audits and qualitative feedback, we identified output anchoring as a key insight: strict safety, brevity, and formatting constraints frequently caused both agents to fall back into a uniform, generic supportive register.
The complete methodology, findings, and design implications for privacy-aware in-vehicle agents were compiled into our research paper, which has been officially accepted at the 18th International Conference on Automotive User Interfaces and Interactive Vehicular Applications (AutomotiveUI '26). Link to the publication under “Related Work.”