Join us for a talk by Eric Schulz, Director of the Helmholtz Institute for Human-Centered AI. This talk is part of the Kempner Seminar Series, a research-level seminar series on recent advances in the field.
Large language models are transforming society while permeating into many applications. I will first propose to use tools from the behavioral sciences to study and better understand these models. I will showcase the utility of this approach by showing how current LLMs behave across a variety of tasks, including learning and decision making, economic, and neuroscientific paradigms, as well as how one can make them more human-like by fine-tuning on psychological data directly. This, in turn, opens up a path towards a cognitive foundation model more generally. I will showcase one such model that has been fine-tuned on 200 canonical psychology tasks, predicts human behavior better than domain-specific model and can generate human-like data for completely new tasks altogether. I will also show early results from another model that maps from behavioral data directly to Python code of the underlying cognitive model.