Reproduce “Agentic Hardware Design as Repository-Level Code Evolution” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Cunxi Yu, Chenhui Deng, Nathaniel Pinckney, Brucek Khailany
The HORIZON framework, developed by Yu et al. at NVIDIA Research, models hardware design as repository-level code evolution, enabling autonomous agents to achieve 100% completion across various RTL benchmark suites through iterative refinement and automated feedback. This approach frames design problems as long-horizon, iterative processes managed within a version-controlled system.
Key insights
- The ability of an agent to effectively iterate and refine designs can lead to full convergence, even starting from low initial pass rates on complex tasks.
- Achieving full correctness on challenging benchmark categories requires a disproportionately high computational budget, indicating that token efficiency is a critical bottleneck for practical application.
- Current benchmarks may not fully capture real-world hardware design complexities, prompting a need for new benchmarks that incorporate robust scoring, randomized tests, and independent reference models to avoid "reward hacking."
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.