Agentic Interpretation: Lattice-Structured Evidence for LLM-Based Program Analysis

arXiv:2605.12694v1 Announce Type: cross Abstract: Large language models can consult information that fixed static analyzers cannot, such as documentation, current security advisories, version-specific metadata, and informal API contracts. This makes LLMs a compelling option for program analyses that depend on information beyond the source program, or that are otherwise not amenable to conventional static analyzers. However, directly asking an LLM for a one-shot whole-program analysis is brittle