Meta Researchers Show AI Agents Can Verify Code Without Running It — and Hit 93% Accuracy

Meta researchers introduced a technique called semi-formal reasoning that enables AI agents to verify code correctness without executing it. By requiring agents to construct explicit premises, trace execution paths, and derive formal conclusions, the approach achieved 93% accuracy on real-world patch equivalence verification — up from 78% with standard reasoning. The technique also improved code question answering by 9 percentage points on RubberDuckBench and fault localization by 5–12 points on Defects4J. The key implication for DevOps teams is that this accuracy level approaches the threshold needed for execution-free reinforcement learning reward signals, potentially reducing the cost and latency of sandbox-based training pipelines. It also offers a middle ground between shallow pattern-matching and expensive formal verification for code review workflows.