Advancing Methodology for Artificial Intelligence for Math Reasoning and Education 47 views

Mathematical reasoning and education are two interrelated and critical application areas of artificial intelligence in the form of Large Language Models (LLMs). The former is well-studied, but it is unclear how and where modern LLMs encode math reasoning capabilities in their parametric knowledge, which could inform interventions to improve performance without catastrophic forgetting on other unrelated tasks. The latter is under-studied, especially as it relates to generating customized, educationally appropriate practice problems and learning materials for students, which has much promise in both advancing the quality of education for learners from all backgrounds and ability levels and reducing the burden on teachers for curating customized learning materials. To address these gaps, this dissertation simultaneously advances methodology around identifying math-specific parameters in LLMs and aligning LLMs for customized math practice problem generation.

PHD (Doctor of Philosophy)

education; large language models; math word problem generation; interpretability and analysis of AI models; math reasoning

English

2025-10-25