Uncertainty-Guided Hierarchical Multi-Agent Planning for Code Inspection and Debugging

ACL ARR 2026 January Submission4175 Authors

Keywords: multi-agent planning, software engineering, hierarchical architecture, uncertainty, LLM

Abstract: The rise of large language models and their utilization in generating or modifying code is now considered common. AI agents now elevate this concept with autonomous and environment-aware decision making for complex code generation, debugging, refactoring, and testing. However, this convenient innovation in is still in its early years, and will require extensive research advancement before it becomes an optimal solution for practical generative software development. In this paper, we propose a multi-agent framework for autonomous code inspection and debugging of AI generated code through an uncertainty-aware hierarchical multi-agent structure. This work focuses on reducing logical and syntax errors in generated code through uncertainty-awareness, and preventive cascading error propagation in large language models caused by confidence blindness. Retrieval augmented generation is used to provide knowledge context from established text-to-code datasets. Proof-of-concept experiments show a pass rate (Pass@1) of 82.28% with a average final uncertainty of 0.0252 and average 0.4731 CodeBleu Score.

Paper Type: Short

Research Area: AI/LLM Agents

Research Area Keywords: Language Modeling, Generation

Contribution Types: Model analysis & interpretability

Languages Studied: Python, English

Submission Number: 4175