Reproduce “FEA-Bench: A Benchmark for Evaluating Repository-Level Code Generation for Feature Implementation” 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
Wei Li, Xin Zhang, Zhongxin Guo, Shaoguang Mao, Wen Luo, Guangyue Peng, Yangyu Huang, Houfeng Wang and 1 more
Researchers from Peking University and Microsoft Research Asia introduced FEA-Bench, a benchmark designed to assess large language models' capability for repository-level incremental code development for new feature implementation. Evaluations on FEA-Bench reveal that current LLMs achieve a resolved ratio below 10%, indicating a significant challenge in this real-world software engineering task.
Key insights
- Current large language models exhibit low performance on repository-level new feature implementation, with the best-performing model achieving a resolved ratio of less than 10%.
- LLMs struggle to effectively leverage excessively long contexts, with increased context length sometimes leading to decreased performance, suggesting that retrieval precision is more critical than mere context size.
- Generating code edits in a 'Natural' format (before-and-after code snippets) significantly outperforms directly generating 'Patch' formatted diffs, due to challenges LLMs face in adhering to strict patch formatting requirements.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.