LARGER boosts code agent accuracy by 13.9 points on locator benchmarks

Researchers have introduced LARGER (Lexically Anchored Repository Graph Exploration and Retrieval), a novel framework designed to improve how coding agents locate files and symbols in large code repositories. The key insight is that existing agents rely heavily on lexical search, missing crucial structural relationships like imports, call chains, type hierarchies, and code-test links. LARGER formalizes this problem as 'Lexically Anchored Structural Localization' and solves it by turning lexical matches into high-precision structural entry points. The framework then performs confidence-filtered local neighborhood expansion within the agent's existing search loop, all without requiring external graph databases or specialized interfaces.

In extensive benchmarking across four tasks—localization, test generation, and codebase understanding—LARGER delivered significant gains. On LocBench, it improved file-level Acc@5 by +13.9 points with tuned hyperparameters and +11.8 points with fixed hyperparameters over the strongest baseline. It also showed consistent improvements on MuLocBench, SWE-Atlas Test Writing, and SWE-Atlas Codebase QA. Because LARGER integrates directly into existing CLI coding agents, it offers a practical, drop-in solution for boosting retrieval accuracy without overhauling agent architectures or adding external tools.