Seminar Data-Science II (Empirical Studies)

The achievement of the learning goals is pursued by practicing on the basis of personally assigned in-depth scientific topics as well as by actively participating in the presentation dates. The organizer will choose subject areas within the field of Data-Science (see Topics) and provide scientific papers to students to work through.

Previous participation in the courses offered by our chair are recommended.

This seminar will be split into two main topic blocks. Every student will be assigned a research paper from only one of these blocks to work on. Yet, it is expected that students also actively participate in discussion on papers from the other topic blocks after they have been presented.

When applying for this seminar, please indicate whether you would be interested in only one or both topic blocks. The two topics we are going to discuss in the HWS 2026 are:

1. LLM Personalization. Large language models (LLMs) increasingly tailor their responses to individual users, a process known as personalization. This can happen explicitly, when a user states preferences or identity, or implicitly, when a model infers characteristics such as gender, ethnicity, or age from subtle cues in how someone writes. While personalization can make systems more helpful, it also raises pressing questions of bias and fairness: models may treat users differently based on inferred group membership, sometimes providing lower-quality or stereotyped responses to members of minority groups. In this topic, we explore how personalization arises in LLMs, how researchers measure the associated biases, and what interventions might mitigate them.
   • Assessing Dialect Fairness and Robustness of Large Language Models in Reasoning Tasks
   • Reading Between the Prompts: How Stereotypes Shape LLM's Implicit Personalization
   • One Persona, Many Cues, Different Results: How Sociodemographic Cues Impact LLM Personalization
2. Sycophancy. Large language models exhibit a tendency known as sycophancy, where they align with a user's viewpoint, even if it is factually incorrect, often to appear more favorable or helpful. This behavior poses a significant risk to the reliability of information and can reinforce user biases. Current research focuses on quantifying this phenomenon and developing mitigation techniques, such as fine-tuning models on synthetic data to distinguish between a user's opinion and factual accuracy. Addressing sycophancy is a critical challenge for ensuring the development of trustworthy and robust AI systems.
   • Chaos with Keywords: Exposing Large Language Models Sycophancy to Misleading Keywords and Evaluating Defense Strategies
   • Measuring Sycophancy of Language Models in Multi-turn Dialogues
   • SycEval: Evaluating LLM Sycophancy
   • When Large Language Models contradict humans? Large Language Models’ Sycophantic Behaviour

Through this seminar, students will gain a comprehensive understanding of the ethical and social dimensions of LLMs, preparing them to critically engage with these technologies in their future work.

On the basis of suitable literature, in particular original scientific articles, students independently familiarize themselves with a topic in data-science, classify and narrow down the topic appropriately and develop a critical evaluation. Students work out concepts, procedures and results of a given topic clearly and with appropriate formalisms in a timely manner and to a defined extent in depth in writing; Evidence of independent development by presenting self-selected examples. Descriptive oral presentation of an in-depth data science topic using suitable media and examples in a given format.

If you are interested in this seminar, please apply to Jana Jung via email. 

Please provide a short motivation to take this seminar, a CV and your Transcript of Records. Also, make sure to indicate which ones of the two given topics you are interested in.