Agentic Interpretation: LLM Program Analysis Framework
A new framework called agentic interpretation proposes combining lattice-based static analysis with LLM-driven program reasoning. The approach decomposes high-level analysis goals into localized claims, tracking the LLM's judgments to expose supported or disputed conclusions. This addresses the brittleness of one-shot whole-program analysis by LLMs, which compress many evidence-dependent judgments into a single opaque answer. The framework aims to leverage LLMs' ability to consult information beyond source programs, such as documentation, security advisories, version-specific metadata, and informal API contracts, while maintaining the discipline of static analysis.
Key facts
- Agentic interpretation combines lattice-based static analysis with LLM-driven program reasoning.
- It decomposes high-level analysis goals into localized claims.
- It tracks the LLM's judgments to expose supported or disputed conclusions.
- One-shot whole-program analysis by LLMs is brittle.
- LLMs can consult documentation, security advisories, version-specific metadata, and informal API contracts.
- The framework brings discipline of lattice-based static analysis to LLM reasoning.
- It addresses the compression of evidence-dependent judgments into a single opaque answer.
- The approach uses intermediate findings to guide later, more focused searches.
Entities
—