Zhao, Z. (CSE) – TOWARD VERIFIABLE REASONING IN LLMS

Chain-of-thought (CoT) prompting can improve final-answer performance, but it does not guarantee that intermediate reasoning steps are faithful, valid, or checkable. This proposal studies how formal methods can make natural-language reasoning more reliable by translating CoT rationales into Lean artifacts, checking the resulting theorem statements and proofs, and using compiler feedback to diagnose and repair failures. The completed work evaluates direct zero-shot and few-shot auto-formalization pipelines for quantity- and logic-focused reasoning problems, measuring proof type-check rate, theorem-statement validity, assumption faithfulness, and repair behavior. The ongoing work extends this pipeline with AMR-guided semantic representations and altered-rationale stress tests. The planned work proposes methods of modeling LLM Agent thinking in Lean. Together, these components separate two questions that are often conflated: whether a model translated the reasoning into the right formal goal, and whether that goal can be proved once translated. The research goal is to develop an evaluation framework and a tool-supported workflow to improve the reliability, auditability, and semantic faithfulness of LLM reasoning.

Event Host: Zekun Zhao, Ph.D. Student, Computer Science & Engineering

Advisor: Jeffrey Flanigan

Zoom: https://ucsc.zoom.us/j/96068207641?pwd=anTpLhBXhIdaIhXKDTB32HlDMA6uIO.1

Passcode: 656037