Qingzhuo Wang

Qingzhuo Wang

M.S. student at Tongji University. Interested in LLM post-training, quantitative research, and agentic RL.

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

Published in arXiv 2026, 2026

Qingzhuo Wang, CaiYi Wang, Jinglu Meng, Ruiyang Qin, Kunyu Peng, Zhihua Wei, Wen Shen

Authorization Closure Graph (ACG) represents user authorization and its dependencies as an evolving, versioned state for tool-using LLM agents. When instructions change, ACG invalidates affected permissions while retaining unaffected authorization. It then computes a minimal repair identifying the missing evidence or authority required for execution, reducing stale-authority errors and unnecessary authorization requests.

Evaluations across three LLMs and two tasks show improved action safety and task success rates.

Code

Recommended citation: Qingzhuo Wang, CaiYi Wang, Jinglu Meng, Ruiyang Qin, Kunyu Peng, Zhihua Wei, Wen Shen. (2026). "Authorization Closure Graph: Minimal Repair for LLM Agents with Evolving User Instructions." arXiv 2026.
Download Paper