arXiv cs.AIMonday · May 25, 2026FREE

Inductive Deductive Synthesis: Enabling AI to Generate Formally Verified Systems

A new paper on arXiv presents Inductive Deductive Synthesis (IDS), the first effective approach to generating formally verified distributed systems using AI. Current coding agents like Codex with GPT-5.4 and Claude Code with Opus 4.6 succeed on only 2 out of 7 distributed key-value-store specifications. IDS, built as an agentic LLM system, jointly and incrementally synthesizes implementation and proof, learning from failed attempts to systematically try promising strategies. It achieves 7/7 specifications in about 6.8 hours and $106 per specification on average, roughly 200x faster than expert effort and 17% cheaper than state-of-the-art agents. IDS also incorporates performance feedback, yielding implementations up to 3x faster. The paper is available on arXiv under ID 2605.23109.