LARGER: Graph-Based Repository Navigation for Coding Agents
A newly introduced framework, known as Lexically Anchored Structural Localization, tackles the challenge of how coding agents at the repository level identify pertinent files and symbols. Presently, these agents depend on lexical searches, overlooking important structural connections such as imports, call chains, type hierarchies, and links between code and tests. While graph-based retrieval can uncover these dependencies, it typically necessitates additional tools that disrupt the agent's workflow. The innovative LARGER (Lexically Anchored Repository Graph Exploration and Retrieval) technique transforms lexical matches into precise structural entry points and reveals confidence-filtered local neighborhoods within the agent's current search process. This strategy seeks to avert localization errors that could impact subsequent tasks like patch creation, test development, and answering questions about the codebase. This research is detailed in arXiv:2605.16352v1.
Key facts
- Repository-level coding agents must localize files and symbols relevant to a task.
- Failures in localization cascade across patch generation, test writing, and codebase QA.
- Existing agents navigate primarily through lexical search.
- Lexical search often misses structural relations like imports, call chains, type hierarchies, and code-test links.
- Graph-based retrieval can recover structural dependencies.
- Existing graph approaches often require separate tools or traversal stages.
- LARGER formalizes repository context localization as Lexically Anchored Structural Localization.
- LARGER turns lexical matches into high-precision structural entry points.
- LARGER exposes confidence-filtered local neighborhoods within the agent's search loop.
- The paper is available as arXiv:2605.16352v1.
Entities
Institutions
- arXiv