New benchmark exposes gaps in AI theorem provers when tackling real-world software verification

Language models have made impressive strides in automated theorem proving within the Lean proof assistant, but progress has been measured almost entirely against mathematical libraries like Mathlib. A new benchmark reveals significant blind spots: LLMs perform far worse when confronted with the kind of complex, interconnected codebases that dominate real software verification work.

The repository-scale problem#

Researchers have released VeriSoftBench, a collection of 500 proof obligations in Lean 4 extracted from actual open-source formal methods projects. Unlike existing benchmarks that isolate individual proofs, these challenges preserve the full repository context—including cross-file dependencies and project-specific libraries—that a real prover would encounter. This shift from curated mathematics to working codebases reveals a crucial gap: what works well on Mathlib transfers poorly to this messier, more realistic setting.

Key findings from LLM evaluation#

Testing both frontier large language models and specialized theorem provers uncovered three patterns. Most strikingly, systems trained specifically for Mathlib-style mathematical proofs lose effectiveness when confronted with software verification. Performance drops dramatically for proofs that depend on large chains of intermediate lemmas and definitions—what researchers call "transitive repository dependence." A proof that traces through many layers of dependencies is substantially harder to automate than one with fewer prerequisites.

Interestingly, limiting the context fed to a prover to only the essential dependency closure—rather than exposing it to the entire repository—actually improves success rates. Yet even with this optimization, significant room remains for improvement.

Why it matters#

The gap between mathematics-focused benchmarks and real-world verification represents a blind spot in progress measurement. To advance automated reasoning for practical software engineering, benchmarks must reflect the actual structure and complexity of the codebases that engineers maintain. VeriSoftBench and its evaluation suite are now available for the community, providing a more grounded yardstick for measuring genuine capability in automated verification.