On Spurious Associations and Large Language Model Alignment

Learn about the complex relationship between LLM alignment goals and unintended output changes in this 45-minute lecture from Victor Veitch of the University of Chicago at the Simons Institute. Explore how alignment procedures designed to make language models more helpful, factual, and polite can lead to unexpected side effects like increased response length. Examine whether these secondary changes are necessary consequences of alignment objectives or merely spurious correlations, drawing from recent research findings in domain adaptation and related areas. Delve into the theoretical and practical implications of distinguishing between essential alignment outcomes and incidental correlations in language model behavior.