Reproduce “SweRank+: Multilingual, Multi-Turn Code Ranking for Software Issue Localization” 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

Revanth Gangi Reddy, Ye Liu, Wenting Zhao, JaeHyeok Doo, Tarun Suresh, Daniel Lee, Caiming Xiong, Yingbo Zhou and 2 more

Researchers from UIUC and Salesforce AI Research developed SWERANK+, a framework for localizing software issues in multilingual codebases. It combines a retrieve-and-rerank model with a multi-turn agent, demonstrating improved accuracy for complex and cross-language bugs.

Key insights

  • Multilingual training not only enables cross-lingual localization but also enhances performance on monolingual tasks, suggesting improved generalization from diverse data.
  • Iterative, multi-turn reasoning is crucial for addressing complex issues, outperforming single-pass methods, particularly when issues have low lexical overlap or involve multiple target functions.
  • A hybrid agentic design, leveraging an efficient retrieve-and-rerank tool with a lightweight LLM agent, provides a practical and effective solution for complex localization without the high cost of pure-agentic systems.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.