LLMs for automatic generation of Dafny helper assertions

Determine whether large language models can automatically generate helper assertions for Dafny programs, producing assertion statements that enable the Dafny verifier to discharge failing verification conditions without manual intervention.

Background

Dafny proofs often require helper assertions to guide the SMT solver through complex reasoning steps, but crafting these assertions is tedious and error-prone for engineers. The paper introduces Laurel, a tool that uses LLMs to generate such assertions, and evaluates techniques like assertion placeholders and example selection to improve success rates.

The conjecture addresses the fundamental feasibility of employing LLMs to synthesize helper assertions in real-world Dafny codebases, a capability that, if established, could reduce brittleness and manual burden in automated verification.