Reproduce “TICoder: A Repository-Level Code Generation Framework with Test-Driven Planning and Implementation-Aware Reuse” 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
Siyu Nan, Yaling Luo, Jian Wang, Neng Zhang, Bing Li
TICoder, developed by researchers at Wuhan University, introduces a repository-level code generation framework that leverages test-driven iterative planning and implementation-aware code reuse to improve accuracy. The framework achieved an average performance increase of 11.25% compared to existing baselines across multiple benchmarks and Large Language Models.
Key insights
- Integrating test cases as behavioral specifications during the planning phase significantly improves the quality and reliability of generated implementation steps by addressing ambiguities in natural language requirements.
- Combining both functional (semantic) and implementation-level similarity for code retrieval is crucial for accurately identifying reusable functions and their appropriate context within a repository.
- Selecting relevant and non-redundant usage patterns for retrieved functions, using perplexity-based filtering, effectively guides LLMs in correctly invoking existing repository code while managing context window limitations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.