ZhuLong: Execution-Grounded LLM Agent Boosts EDA Scripting Accuracy
A new coding agent named ZhuLong has been developed by researchers to tackle the challenges posed by undocumented APIs in EDA scripting, particularly focusing on tool-specific long-tail issues. This innovative system, outlined in a paper on arXiv (2608.07925), is aimed at PyAether and SKILL languages, utilizing a combination of API retrieval, documentation analysis, and sandbox execution through unified MCP tools. A notable feature is its offline API self-exploration, which deduces undocumented behaviors via counterfactual experimentation. ZhuLong was tested on EDA-Eval-PyAether, a benchmark comprising 158 real-world tasks, achieving a 78.5% Pass@1 in the Empyrean Aether environment, significantly surpassing the 23.6% of a pure LLM baseline. The study, authored by researchers from the University of Texas at Austin and Empyrean Technology, Inc., was released on August 1, 2025.
Key facts
- ZhuLong is an execution-grounded LLM coding agent for EDA scripting.
- It targets PyAether and SKILL scripting languages.
- It combines API retrieval, documentation inspection, and sandbox execution via unified MCP tools.
- It features an offline API self-exploration mechanism using counterfactual experimentation.
- Evaluated on EDA-Eval-PyAether benchmark with 158 real-world tasks.
- Achieved 78.5% Pass@1 in the commercial Empyrean Aether environment.
- Pure LLM baseline achieved only 23.6% Pass@1.
- Sandbox execution removal caused a 41.2 percentage point drop in performance.
- Self-exploration mechanism added 3.2 percentage points accuracy and reduced tool calls by 22.1%.
- Paper available on arXiv with ID 2608.07925.
Entities
Institutions
- University of Texas at Austin
- Empyrean Technology, Inc.
- arXiv
Locations
- Austin
- United States