Recoleta
GitHub
Source note

Retrieval-Oriented Code Representations in Agentic Bug Localization

Authors
Genevieve Caumartin, Tse-Hsun, Chen, Diego Elias Costa
Source
arXiv
Published
Collection
Software Intelligence
Code RetrievalBug LocalizationSoftware AgentsCode RepresentationLLM Ranking
Open arXivMarkdown source

Summary

The paper studies how the textual representation of repository files affects file-level bug localization for software agents. Role-aware summaries usually provide the best accuracy-cost balance, while combining representations and LLM reranking improves coverage.

Problem

  • Repository-scale bug localization must identify relevant files among hundreds or thousands of candidates before an agent can generate or validate a patch.
  • Raw source code can exceed context limits and increase token and retrieval costs, while file paths omit useful semantic information.
  • The study measures how representation choice affects retrieval accuracy, ranking quality, and representation footprint on Long Code Arena and SWE-bench Verified.

Approach

  • Compare five file representations: file paths, raw source code, role-aware summaries, detailed technical summaries, and synthetic bug-report summaries.
  • Retrieve candidate files with BM25, dense embeddings, or LLMs, then use reciprocal rank fusion or an LLM to rerank candidates.
  • Measure localization with MAP and Hit@k, while estimating cost through token-based representation footprint and summary-generation API cost.
  • Test the strongest methods inside the Agentless file-localization pipeline.

Results

  • Role-aware summaries improve Hit@5 over file-path representations by up to 40% while using a representation footprint 10.4 to 20.9 times smaller than raw source code.
  • Reciprocal rank fusion of complementary LLM retrieval results reaches Hit@5 of 89.3% on Long Code Arena and 83.4% on SWE-bench Verified.
  • Combining representations improves localization by up to 31.9%, and LLM post-retrieval ranking improves results by up to 42.0%.
  • In the Agentless case study, role-aware summaries with LLM ranking reach 94% Hit@6, a 4.7 percentage-point gain over the baseline.
  • The paper reports no single representation that wins across every model, dataset, and pipeline stage; role-aware summaries provide the strongest overall cost-effectiveness trade-off, while raw source code can perform well at much higher cost.

Link