Problem Description

Large Language Models (LLMs) are increasingly used as conversational agents in experiments—for example, as customer-service chatbots, digital advisors, or social interaction partners. Compared to traditional experimental stimuli, however, LLM-based agents are difficult to control: small changes in prompts, conversation context, or model parameters can result in substantially different responses.

Moreover, different models may behave differently even with identical instructions. Proprietary models accessed through APIs can be silently updated, changed, or discontinued by their providers. If the original model version is no longer available, an experiment cannot be replicated exactly—even when the prompt and experimental procedure have been documented.

This raises a fundamental methodological question: To what extent can LLM-based agents be considered reliable and reproducible experimental stimuli? The topic is increasingly discussed across NLP, HCI, Information Systems, and Computational Social Science, but practical implications for experiments with human participants remain unclear.

Goal of the Seminar Paper

The student should explore and critically discuss the challenges of using LLM-based conversational agents in experimental research.

Possible aspects include:

• How prompt sensitivity, stochasticity, model choice, and model updates affect agent behavior
• Differences between open-weight and proprietary LLMs with regard to transparency and replicability
• Consequences for experimental validity, comparability, and reproducibility
• Existing recommendations for documenting prompts, parameters, model versions, and access dates
• Practical ways to make future experiments with LLM agents more robust and transparent

The paper should be based on a broad literature research and develop a critical perspective on the topic. It should conclude with initial recommendations for researchers who plan to use LLM-based conversational agents in experiments.

Requirements

• Interest in LLMs, conversational agents, and experimental research
• Ability to read and synthesize interdisciplinary literature
• Basic understanding of experimental research is beneficial

Research Method

Systematic Literature Review (SLR), complemented by a conceptual-comparative synthesis along the analytical dimensions outlined above.

Starting Points

• Chen, L., Zaharia, M., & Zou, J. (2023). How Is ChatGPT’s Behavior Changing over Time? arXiv:2307.09009. https://arxiv.org/abs/2307.09009
• Sclar, M. et al. (2024). ProSA: Assessing and Understanding the Prompt Sensitivity of LLMs. arXiv:2410.12405. https://arxiv.org/abs/2410.12405
• Bender, E. M. et al. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? FAccT. https://doi.org/10.1145/3442188.3445922
• Aher, G. V., Arriaga, R. I., & Kalai, A. T. (2023). Using Large Language Models to Simulate Multiple Humans and Replicate Human Subject Studies. ICML. https://proceedings.mlr.press/v202/aher23a.html