Almost Sure Reasoning: Generating Verified Formalizations with Language Models and Logical Solvers

Mohammad Raza, Natasa Milic-Frayling

Keywords: Automated Reasoning, SAT/SMT Solvers, Formal Methods, Large Language Models

TL;DR: A novel approach to combining language models with logical solvers that advances accuracy of reasoning and also provides a verification mechanism with near-perfect precision.

Abstract: Robustness of reasoning remains a challenging problem for large language models, and addressing it is crucial for advancing the reliability and practical application of AI-driven reasoning systems. We introduce Semantic Self-Verification (SSV), a novel approach that addresses the key challenge in combining language models with the rigor of logical solvers: to accurately translate the reasoning problem from natural language to the formal language of the solver. SSV produces strong abstract formalizations of problems by verifying and refining them against concrete instantiations that are generated by the model and verified by the solver. In addition to significantly advancing the overall reasoning accuracy over the state-of-the-art, a key novelty that this approach presents is a feature of verification that has near-perfect precision over a significant coverage of cases, as we demonstrate on open reasoning benchmarks. We propose such $\textit{near-certain reasoning}$ as a new approach that can reduce the need for manual human verification in many cases, taking us closer to more dependable and autonomous AI reasoning systems.

Primary Area: neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 2446