A new method, Inductive Deductive Synthesis (IDS), allows AI agents to jointly synthesize implementations and formal proofs, dramatically accelerating the generation of formally verified systems compared to expert human effort.
AI agents excel at generating code but struggle with tasks requiring full formal guarantees, particularly in complex distributed systems. This paper introduces Inductive Deductive Synthesis (IDS), an agentic LLM system designed to address this gap by jointly synthesizing implementation and proof, learning from failures, and systematically exploring promising strategies.
IDS achieved 7/7 success on distributed key-value-store specifications in just 6.8 hours, demonstrating performance roughly 200 times faster than expert effort and 17% cheaper than state-of-the-art coding agents. Furthermore, by incorporating performance feedback, IDS yielded implementations up to 3x faster than previously published verified systems, pushing the boundaries of AI-assisted formal verification.