OpenROAD Agent: An Intelligent Self-Correcting Script Generator for OpenROAD

Authors

B.-Y. Wu, U. Sharma, A. Rovinski, and V. A. Chhabria

Abstract

Large language models (LLMs) are increasingly being used in various domains, including chip design. Recent works have demonstrated the effectiveness of LLMs in EDA tool script generation. However, these LLMs can often hallucinate API calls even when directly provided API documentation and context. As a result, this leads to significant errors and inefficiencies in chip design workflows. To address this, we introduce OpenROAD Agent, an LLM that integrates directly with OpenROAD, an open-source tool for physical design. OpenROAD Agent enables real-time script generation with error detection and correction. OpenROAD Agent autonomously generates and executes Python code within OpenROAD while dynamically refining outputs based on tool feedback. OpenROAD Agent leverages Qwen2.5-Coder and employs a hybrid supervised and reinforcement learning training to improve code correction capabilities, usability, and hallucination frequency. OpenROAD Agent achieves an accuracy of 94% on script generation tasks and outperforms both prior work and existing foundation models.