Authorization Closure Graph: Minimal Repair for LLM Agents with Evolving User Instructions

2026-09-26

AI Summary

This study addresses the problem that large language model (LLM) agents lack selective authorization update mechanisms when user instructions change, leading to permission invalidation or redundant requests. To this end, we propose the Authorization Closure Graph framework, which pioneers graph-based authorization dependency modeling and state versioning. By leveraging a dependency analysis algorithm, the framework precisely identifies nodes affected by instruction revisions and computes minimal repair solutions for local permission updates, thereby avoiding full re-authorization. Experiments across three mainstream LLMs and two tasks demonstrate that our approach significantly improves both action safety rates and task success rates. The source code has been made publicly available.

Abstract

Tool-using large language model (LLM) agents increasingly perform state-changing actions that require user authorization. Yet existing approaches do not provide a principled mechanism for selectively updating prior authorization when only part of an instruction changes. To this end, we propose an Authorization-Closure-Graph (ACG)-based framework that represents authorization and its dependencies as an evolving, versioned state. ACG selectively invalidates authority affected by a revision while preserving unaffected portions of the authorization state, and computes a minimal repair that identifies only the missing evidence or authority required for execution. This enables agents to adapt to revised instructions while avoiding stale authority and unnecessary authorization requests. We evaluate ACG across three advanced LLMs in two natural tasks, and ACG consistently improves action safety rate and task success rate. Code is available at https://github.com/weiliang822/ACG.

Problem

• LLM agents
• user authorization
• evolving instructions
• minimal repair
• tool use

Innovation

• Authorization Closure Graph
• Minimal Repair
• LLM Agents
• Evolving Instructions
• Tool Authorization