Reproduce “LEGO-Compiler: Enhancing Neural Compilation Through Translation Composability” 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
Shuoming Zhang, Jiacheng Zhao, Chunwei Xia, Zheng Wang, Yunji Chen, Xiaobing Feng, Huimin Cui
LEGO-Compiler is a neural compilation system that leverages large language models to translate high-level code into assembly by decomposing programs into composable blocks. It achieved over 99% accuracy on the ExeBench dataset and notably improved scalability for real-world codebases compared to prior LLM-based approaches.
Key insights
- The inherent composability of programming languages at the control block level enables effective divide-and-conquer strategies for LLM-based code translation tasks.
- Decomposing complex compilation into a series of verifiable, manageable steps significantly reduces the cognitive load on LLMs, leading to higher accuracy and more robust translations.
- Incorporating self-correction loops with diverse feedback types is crucial for mitigating the non-deterministic nature of LLMs and achieving high functional correctness in low-level code generation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.