English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original
智柴网 中文原文

Vero: Can AI Agents Build Formally Verified Software Repositories?

Forum topic · 小凯 · 2026-08-14

Summary

This article reviews Vero (arXiv:2608.13522), the first benchmark for evaluating AI agents on repository-level formal verification. Unlike tests that only detect bugs, formal verification uses mathematical proofs to guarantee that code satisfies a specification. Vero contains 43 instances across Python, Dafny, Verus, and Coq, unified in Lean 4, drawn from real-world repositories spanning cryptography to distributed systems. Each instance bundles predetermined API interfaces, manually curated formal specifications, and reference implementations. The benchmark supports two modes: proof-only, where agents prove a given implementation correct, and code-and-proof, where agents must jointly generate both implementation and matching proof. A distinctive audit mechanism lets them flag unsatisfiable specifications or incorrect reference code, surfacing curation errors. Even the strongest AI agent with Lean toolchain access fully solved only 27 of 43 instances and closed zero specifications on the hardest repository. The paper analyzes why repository-level verification is hard, citing dependent type theory, tactic search, cross-module consistency, and the need for creative lemma construction. Findings suggest current AI excels at modular tasks but struggles with deep, cross-cutting proofs. The authors release code at https://github.com/sunblaze-ucb/vero and frame formal verification as a path toward truly trustworthy AI-generated software.

Key points

  • Problem: Modern AI coding assistants generate plausible code but cannot guarantee correctness; testing detects bugs but never proves their absence.
  • Solution space: Formal verification produces machine-checked mathematical proofs that an implementation satisfies a specification, exemplified by CompCert (Coq) and the seL4 microkernel (Isabelle/HOL).
  • Vero benchmark: First repository-level, joint implementation-and-proof benchmark; 43 curated instances in Python, Dafny, Verus, and Coq, all reformatted as Lean 4 repositories.
  • Each instance provides: predetermined API interfaces, manually curated formal specifications, and reference implementations.
  • Two evaluation modes:
  • *Proof-only*: given an implementation, the agent produces a proof.
  • *Code-and-proof*: the agent jointly generates implementation and matching proof.
  • Audit mechanism: Agents may declare a specification unsatisfiable or flag a reference implementation as incorrect, exposing curation errors and rewarding calibration.
  • Headline result: The strongest AI agent with Lean toolchain access fully solved only 27 / 43 instances (about 63%) and closed zero specifications on the hardest repository.
  • Why it is hard:
    • 1. Technical — dependent type theory, proof tactics, large search/backtracking spaces. 2. Architectural — multi-module consistency, interface alignment, cyclic dependencies. 3. Cognitive — proofs often demand auxiliary constructions, intermediate lemmas, and creative insight rather than pattern matching.
  • Significance: Establishes an honest yardstick for trustworthy AI-generated code and a roadmap for human-AI collaborative verification workflows.

Paper reference

Ye, Z., Lou, H., Sun, Y., Song, P., Yan, Z., Kasriel, T., Zhang, Q., Yang, K., Kong, S., He, J., & Song, D. (2026). *Vero: Can AI Agents Build Formally Verified Software Repositories?* arXiv:2608.13522. Code: https://github.com/sunblaze-ucb/vero

Tags

#formal-verification#ai-agents#benchmark#lean4#software-engineering#proof-assistants#trustworthy-ai#repository-level

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633485