Improving Language Models' Reasoning: A Systematic Perspective

Abstract

This talk discuss how recipes for building LLMs of strong reasoning capability from first principles. We envisage large language models to become the next-generation computational platform and foster an ecosystem of LLM-based new applications. This naturally requires the foundation models to perform complex tasks that often involve the composition of linguistic and logical operations. We first review the generic recipe for building large language models. Then we discuss recipes for improving language models’ reasoning capabilities and the corresponding evaluation. Finally, we consider further improvements by complexity-based prompting, distilling chain-of-thought, and learning from AI feedback.