Semi-formal reasoning helps agents reason about code without executing the code (arxiv.org)

A new study introduces "agentic code reasoning," a method allowing AI agents to analyze and reason about code semantics without executing it. This approach employs "semi-formal reasoning," which requires agents to construct explicit premises and trace execution paths to reach formal conclusions, thereby ensuring that their reasoning is sound and supported. The research evaluated this technique across three critical tasks: patch equivalence verification, fault localization, and code question answering, demonstrating significant accuracy improvements. For instance, it increased accuracy in patch equivalence from 78% to 88%, and even reached 93% for real-world patches. The significance of this advancement lies in its potential applications in areas like reinforcement learning (RL) training, code review, and static program analysis, where execution is often impractical or costly. By enabling more reliable semantic code analysis without execution, semi-formal reasoning could enhance the efficiency and accuracy of AI-driven development processes, making it a valuable tool for the AI/ML community in software engineering.