FriSem

Michael Frank, Professor of Psychology, Stanford University

Title: Large language models as cognitive models? 

Abstract: LLMs are powerful new technologies for natural language processing. Their impressive, seemingly human-like capacities raise questions about what - if anything - they can tell us about human cognition. In this talk, I'll discuss similarities and differences between LLMs and previous cognitive models, with the general thesis that models are useful to the extent that you can perform causal interventions on them that are difficult or impossible to perform on the target system of interest. The key interventions for LLMs include behavioral evaluation, causal probing, and controlled rearing. I'll use case studies including relational learning and language acquisition to highlight opportunities and challenges relating to each of these interventions.