Automating Formal Verification with Reinforcement Learning and Recursive Inference
Researchers have developed novel methods to automate formal verification for large language models, addressing the scarcity of data for proof assistants and verification-aware languages. Their approach utilizes reinforcement learning with verifiable rewards (RLVR) and verifier-guided inference search. Experiments in Dafny showed a significant increase in verified reward, though specification hacking was identified as a challenge. Further refinements and a verifier-guided scaffold in Lean improved proof generation success rates on specific benchmarks. AI
IMPACT Enhances AI's ability to generate provably correct code and proofs, crucial for safety-critical systems.