Meta researchers verify code patches without running them at 93% accuracy

Meta researchers published a technique called semi-formal reasoning that enables AI agents to verify whether two code patches are functionally equivalent without executing either one, hitting 93% accuracy on real-world agent-generated patches compared to 78% with standard reasoning approaches. The method works by forcing agents to construct explicit premises, trace execution paths, and derive formal conclusions rather than making unsupported inferences, which catches the kinds of errors that occur when agents assume function behavior without actually tracing call chains. The practical payoff for AI coding infrastructure is significant: training pipelines that currently require expensive sandbox execution at every evaluation step could shift to structured semantic verification, reducing both cost and latency at scale.