Reproduce “VeriPilot: An LLM-Powered Verilog Debugging Framework” on OpenResearch

About this paper

Yihan Wang, Cheng Liu, Jiazheng Zhang, Lei Zhang, Long Cheng, Xiaowei Li, Huawei Li

VeriPilot is an LLM-powered multi-agent framework designed to automate Verilog debugging by integrating static structural analysis, dynamic simulation, and LLM reasoning with golden reference models for fine-grained internal signal comparison. It achieved repair success rates of 92.98% on the Strider benchmark and 91.43% on the CVDP benchmark using GPT-5, surpassing traditional methods and direct LLM prompting.

Key insights

• Leveraging golden reference models for fine-grained internal signal comparison and cross-abstraction semantic alignment is crucial for overcoming the limitations of coarse-grained debugging feedback.
• Systematic program analysis techniques, such as CDFG-based tracing, can precisely identify the earliest points of behavioral divergence, enabling the localization of minimal suspicious code regions.
• Presenting LLMs with structured debugging evidence, including localized errors and corresponding correct logic from a golden model, significantly improves repair synthesis accuracy and efficiency.

Summary by alphaXiv.