Out-of-Distribution Generalization as Reasoning - Are Large Language Models Competitive?

Explore a thought-provoking lecture on out-of-distribution generalization and its connection to reasoning in large language models. Delve into the concept of humans answering questions about unfamiliar scenarios by chaining together beliefs learned from examples. Examine the robust logic framework and its application to probabilistic soundness in combining beliefs. Investigate the principles of PAC learning and how they relate to the success of sound chaining in making predictions. Consider the importance of modularity in the world for enabling separately learnable rules across different feature sets. Analyze the power of the robust logic framework in the context of out-of-distribution generalization and its relevance to large language models. Join Les Valiant from Harvard University as he presents this insightful talk at the Simons Institute, addressing emerging generalization settings and challenging our understanding of machine learning capabilities.

Out-of-Distribution Generalization as Reasoning: Are LLMs Competitive?