Overview
Explore a comprehensive analysis of large language models' mathematical reasoning capabilities in this 37-minute video by Yannic Kilcher. Dive into the findings of a paper from Apple that challenges the current understanding of LLMs' performance on high-school level math questions. Learn about GSM-Symbolic, a new benchmark designed to provide more reliable metrics and insights into the reasoning abilities of language models. Discover how LLMs exhibit variance in responses to similar questions and how their performance declines when numerical values are altered. Examine the fragility of mathematical reasoning in these models and understand why their capabilities deteriorate as question complexity increases. Gain valuable insights into the limitations of current LLMs in performing genuine logical reasoning and the implications for future development in the field of artificial intelligence and natural language processing.
Syllabus
GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models
Taught by
Yannic Kilcher