Reproduce “A Problem-Oriented Perspective and Anchor Verification for Code Optimization” on OpenResearch

About this paper

Tong Ye, Tengfei Ma, Xuhong Zhang, Hang Yu, Jianwei Yin, Wenhai Wang

This research introduces a problem-oriented perspective for constructing code optimization datasets and an anchor verification framework for Large Language Models (LLMs). This approach enables LLMs to achieve substantially higher code optimization rates and speedups, for instance, a 58.90% optimization rate and 5.22x speedup with QWEN2.5-CODER 32B, while also improving the functional correctness of optimized code.

Key insights

• Problem-oriented datasets exhibit greater structural and semantic diversity (e.g., higher Graph Edit Distance) and promote global algorithmic optimizations more effectively than user-oriented datasets.
• Leveraging the functional correctness of the original slow code as an "anchor" for test case generation is crucial for mitigating the "optimization tax" and improving the reliability of LLM-optimized code.
• The problem-oriented approach significantly improves data efficiency, allowing LLMs trained on a fraction of problem-oriented data to outperform those trained on full user-oriented datasets.

Summary by alphaXiv.