Lang2Logic: Reducing Hallucinations in Logical Translation

Time: 2024
Location: Penn State University
Advisor: Prof. Mahfuza Farooque
Skills: NLP, Supervised Fine-Tuning (SFT), SciPy

To address the unreliability of Large Language Models in formal reasoning, I developed ‘Lang2Logic’, a framework that translates natural language into Conjunctive Normal Form (CNF) for automated theorem proving. I fine tuned LLMs and constrained their outputs using a custom defined grammar and symbolic computation libraries, successfully creating a pipeline that significantly reduces hallucinations and enables the expansion of SAT solver domains from logical verification to finance, healthcare, and other fields.