Reproduce “TraceCoder: A Trace-Driven Multi-Agent Framework for Automated Debugging of LLM-Generated Code” 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
Jiangping Huang, Wenguang Ye, Weisong Sun, Jian Zhang, Mingyue Zhang, Yang Liu
TraceCoder, a multi-agent framework developed by researchers across China and Singapore, automates the debugging of LLM-generated code by utilizing fine-grained runtime traces and a historical lesson learning mechanism. The system consistently achieved the highest Pass@1 accuracy, for instance, reaching 82.00% on the complex ClassEval benchmark with Gemini-2.5-Flash-0417, representing a 34.43% relative improvement over prior methods.
Key insights
- Providing 'white-box' visibility through fine-grained runtime tracing is crucial for precise fault localization and causal analysis in LLM-based code repair.
- Learning from historical debugging lessons and avoiding repetitive mistakes significantly enhances the efficiency and success rate of iterative program repair.
- A collaborative multi-agent architecture can effectively deconstruct and manage the complex, iterative process of automated code debugging.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.