ARTFEED — Contemporary Art Intelligence

Agentic Interpretation: LLM Program Analysis Framework

ai-technology · 2026-05-14

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

Sources