Publication Date
Fall 2025
Degree Type
Master's Project
Degree Name
Master of Science in Computer Science (MSCS)
Department
Computer Science
First Advisor
Christopher Pollett
Second Advisor
Navrati Saxena
Third Advisor
Prasanna Nikhil Sathwik Vadlamani
Keywords
Large Language Models, LLMs, Mathics, Lean, Chain-of-Thought prompting, Deepseek, Post-tuning
Abstract
Large language models (LLMs) are good at natural language tasks but are not up to the mark when it comes to mathematics and theorem-proving, as they rely on language patterns instead of understanding the problem and thinking through the solution. This project addresses issues such as a lack of exposure to structured datasets, difficulty with generating outputs that require multi-step reasoning, and limitations of short context. We fine-tune pre-trained LLMs on structured datasets like MATH, GSM8k, open-r1, deepseek-prover, and OpenBootstrappedTheorem. We integrate two software tools, Mathics and LEAN, and enhance reasoning through Chain-of-Thought (CoT). Additionally, we conduct the experiments using state of the art Mixture of Experts (MoE) and parameter efficient fine-tuning (PEFT) techniques such as LoRA and DoRA. The outcomes of this project are better model performance on complex math problems, and particularly on formal theorem-proving datasets, which is a comparatively understudied domain in recent LLM research. This takes a step toward developing and fine-tuning models that can handle challenging mathematical and logical domains.
Recommended Citation
Rohan Kumar Bayya, Naga, "Enhancing LLM’s Mathematical and Theorem Proving Abilities" (2025). Master's Projects. 1616.
DOI: https://doi.org/10.31979/etd.wdma-vp7y
https://scholarworks.sjsu.edu/etd_projects/1616
