Agentic Interpretation: Lattice-Structured Evidence for LLM-Based Program Analysis
Jacqueline L. Mitchell, Chao Wang
Why It Matters
What makes this one worth your time
This framework could improve program analysis by leveraging LLMs to incorporate dynamic information sources, potentially leading to more accurate and comprehensive analyses.
Agentic interpretation combines lattice-based static analysis with LLMs for enhanced program reasoning.
Summary
The paper proposes a framework called agentic interpretation, which integrates lattice-based static analysis with large language models (LLMs) for program analysis. This approach decomposes high-level analysis goals into localized claims and uses a worklist algorithm to manage the evolution of these claims and their judgments.
Key contributions
- Introduction of agentic interpretation framework for LLM-driven program analysis.
- Formal model of agentic interpretation integrating lattice-based static analysis.
- Illustration of the approach with a worked example involving third-party components.
Notable insights
- The use of a lattice structure to track LLM judgments provides a systematic way to manage evidence-dependent judgments.
- Decomposing analysis goals into localized claims allows for more focused and iterative program reasoning.
Possible limitations
- Not stated in the abstract
Abstract
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 because it compresses many evidence-dependent judgments into a single opaque answer, rather than exposing which conclusions are supported or disputed and using intermediate findings to guide later, more focused searches. In this paper, we propose agentic interpretation, a framework that brings the discipline of lattice-based static analysis to LLM-driven program reasoning. At a high level, agentic interpretation decomposes a high-level analysis goal into localized claims, and tracks the LLM's judgment about each claim in a finite-height lattice. A worklist algorithm governs how claims and their judgments evolve during the analysis. We introduce a formal model of agentic interpretation, explore the design space it opens, and illustrate the approach with a worked example analyzing code that depends on opaque third-party components.