Evident: LLMs assist but don't reason in program analysis

A new paper from Haonan Li, Tianyang Zhou, Manu Sridharan, Hang Zhang, and Zhiyun Qian (UC Riverside and collaborators) argues that LLMs should not directly reason about program behavior for bug analysis. While LLMs are increasingly used to judge whether a potential bug is triggerable, the authors warn that plausible explanations are insufficient for dismissing warnings, especially for 'no-bug' decisions. They propose Evident, a system that restricts the LLM to constructing a warning-specific analysis harness — essentially a scenario for formal checking — while the actual reasoning about reachability and execution context is handled by a backend formal analysis tool.

Evident was evaluated on 200 real warnings from two static detectors for Android kernel drivers. It correctly classified 151 cases (76%), discharging 111 false alarms without letting a single confirmed bug slip through. Notably, Evident also rediscovered a confirmed vulnerability that had been missed by both prior LLM-based filtering and manual triage. The remaining cases were either unresolved or conservatively retained as potential bugs. This hybrid approach demonstrates that separating LLM assistance from formal reasoning can yield high precision while maintaining safety — a crucial advance for deploying AI in critical software analysis.