AI seminar and lunch

​Johannes Gasteiger, research scientist at Google Research, is coming to EPFL for a talk on "Knowledge, Truthfulness, Honesty, and Deception in LLMs". His talk will be followed by a lunch.

​Registration is mandatory, register here: https://go.epfl.ch/ai-lunch

About the talk: Knowledge, Truthfulness, Honesty, and Deception in LLMs

​Abstract: Large language models (LLMs) present capabilities never before seen in ML systems. However, they also present numerous new challenges. Importantly, we only have very rough control and understanding of their outputs and inner workings. Notions such as knowledge, communicative intent, or honesty are critical for this understanding. Unfortunately, these terms are hard to grasp even for regular human communication, and this becomes even worse for human-machine communication. Progress on these topics is critical for controlling LLMs and using them for human benefit ‒ especially as their behavior becomes more agentic and goal-oriented.
​In this talk, I will first give a brief overview of current research on how knowledge is represented in LLMs. I will then distinguish the notions of truthfulness and honesty, and discuss current research in truthfulness and factuality. I will particularly focus on our recent analysis of unsupervised methods for discovering latent knowledge. Finally, I will discuss the topic of honesty, focusing in particular on its most problematic variant: Deception. I will first approach this notion from a theoretical angle and discuss what deception even means in this context and which definitions might be workable for AI systems. Based on these definitions, I will show how prevalent deception already is in current AI systems.

​Bio: Johannes Gasteiger is a research scientist at Google Research in Zurich. His research is focused on the safety, interpretability, and factual groundedness of advanced ML models such as LLMs. During his PhD in Stephan Günnemann's group at TU Munich he studied how to jointly leverage both geometry and structure in GNNs.