AI and the World's Hard Problems · Theoretical computer science · arXiv · 2026-10-23
RAIN: Agentic Proof Generation for Rocq-based Software Verification
Constructing proofs for interactive theorem provers is labor-intensive despite the strong correctness guarantees they offer. The authors present RAIN, a lightweight proof-specialized agent that adds repository exploration, iterative refinement, and checker feedback to an LLM, allowing it to retrieve lemmas and revise whole proof scripts. On a subset of CoqStoq, RAIN achieved over 90% prove success, at least 24.7% higher than existing baselines, and matched general-purpose coding agents at significantly lower cost.