Vero: Can AI Agents Build Formally Verified Software Repositories?

Imagine you're developing an AI that can write complex software, but you need to ensure that the code it generates is correct and meets specific requirements. Currently, many AI systems can produce code, but they often lack a way to verify that the code works as intended, leading to potential errors and vulnerabilities. This is particularly problematic in multi-module systems where the interactions between different parts of the code can introduce unforeseen issues, a situation known as integration failure. Existing benchmarks typically focus on isolated functions or only on generating proofs without considering the actual implementation, which doesn't reflect the real-world challenges developers face.

To address this gap, the authors created Vero, a benchmark designed to evaluate both the implementation of code and the generation of formal proofs at the repository level. Vero includes 43 multi-module instances from actual codebases, covering various programming languages and domains. It allows for a comprehensive assessment of AI agents by providing curated specifications and reference implementations, along with an audit mechanism that helps identify and correct errors in both code and specifications. This means that Vero not only tests how well an AI can generate code but also how effectively it can ensure that the code meets its intended specifications, providing a more realistic measure of progress in verified software synthesis compared to previous benchmarks.

Vero introduces a new benchmark for evaluating joint implementation and proof synthesis in multi-module codebases.

The benchmark includes a robust audit mechanism and diverse real-world instances, enhancing evaluation reliability.