Natural Language Processing (2/2)

Friday 28 Nov 2025, 14:15 → 15:45 Europe/Berlin

DO: JvF25/3-303 | BN: b-it/1.047

Why Making AI Prefer Accordion Polka Remixes Can Make It Evil by Florian Mai

Emergent Misalignment is the phenomenon that fine-tuning Large Language Models on a dataset from a narrow domain can make it become broadly misaligned. In this lecture, we discuss why this matters, why it happens, and what we can do to prevent it.

LLM-based control architectures for autonomous robots by Oliver Urbann

This presentation explores the potential of control architectures derived from Large Language Models (LLMs). This emerging field presents a largely uncertain direction. Consequently, we commence with conventional control architectures and illustrate scenarios where LLMs can augment their capabilities. These include the Subsumption architecture, hierarchical and hybrid representations. Modern approaches such as Behaviour Trees are also addressed. Following a theoretical introduction of LLM application in robotics, we present an example from robot soccer at TU Dortmund University and two contemporary example projects from the Fraunhofer Institute for Material Flow and Logistics. These insights provide practical applications.