Can LLMs Reason Like Automated Theorem Provers for Rust Verification? VCoT-Bench: Evaluating via Verification Chain of Thought

Researchers Zichen Xie and Wenxi Wang have published a groundbreaking paper introducing VCoT-Bench, a new framework that challenges whether Large Language Models (LLMs) can truly reason like automated theorem provers for Rust program verification. The core innovation is VCoT-Lift, which transforms low-level solver reasoning into high-level, human-readable verification steps, creating an explicit Verification Chain-of-Thought. This provides concrete ground truth for evaluating whether models understand the logical deductions required for verifying nontrivial Rust code, moving beyond simple pass/fail assessments.

Using this framework, the researchers built VCoT-Bench—a comprehensive benchmark of 1,988 VCoT completion tasks designed to rigorously test LLMs' understanding of the entire verification process. The benchmark measures performance across three orthogonal dimensions: robustness to varying degrees of missing proofs, competence across different proof types, and sensitivity to proof locations. When evaluating ten state-of-the-art models, the results revealed severe fragility, indicating current LLMs fall well short of the reasoning capabilities exhibited by automated theorem provers.

The findings are particularly significant as LLMs increasingly assist in secure software development. The research demonstrates that while models might generate seemingly correct proof hints, they often lack the deep logical understanding required for reliable Rust verification. This gap between surface-level performance and genuine reasoning capability highlights a critical limitation in current AI systems when applied to safety-critical programming tasks where rigorous verification is essential.