Reproduce “CodeTeam: An LLM-Powered Multi-Agent Framework for Repository-Level Code Generation” 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
Yifei Wang, Ruiyin Li, Peng Liang, Qiong Feng, Zengyang Li, Mojtaba Shahin, Arif Ali Khan
The CodeTeam framework, developed by researchers from Wuhan University and collaborators, introduces an LLM-powered multi-agent system for generating complete software repositories from natural language requirements. The framework improved SketchBLEU scores by 4.1 absolute points and execution-based test pass rates by 4.4 absolute points (PE setting) compared to existing baselines, specifically by reducing structural and organizational errors.
Key insights
- Structured multi-agent workflows, particularly those with explicit, machine-checkable design contracts like SDS, significantly enhance the robustness and correctness of repository-level code generation.
- Upfront architectural planning, including diversified proposal generation and CTO-led selection, plays a crucial role in reducing downstream packaging and import errors during implementation.
- Mechanisms like retrieval-augmented planning and dynamic developer allocation are critical for producing valid initial designs and improving the efficiency and consistency of the multi-agent implementation process.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.