LocalizeAgent
An evidence-grounded agent for localizing design issues in Java programs using structured program facts and iterative reasoning.
LocalizeAgent investigates design-issue localization as an evidence-grounded reasoning task. It combines structured program facts with iterative LLM reasoning so each localization claim can be traced back to concrete code evidence.
Design issues such as misplaced responsibilities and weak cohesion are diffuse and context dependent. Without structured program evidence, LLM localizations can appear plausible while remaining weakly anchored in the code.
LocalizeAgent extracts dependency graphs, call graphs, responsibility profiles, and coupling signals, then uses an iterative agent loop to reason over those facts and refine the suspected fault location.
On real-world Java refactoring data, LocalizeAgent reports relative exact-match accuracy gains of 138%, 166%, and 206% for information hiding, complexity, and modularity issues, respectively.
Evidence-grounded localization provides a stronger basis for downstream refactoring because the agent's recommendation is connected to explicit program facts.
Code is available at github.com/fraolBatole/LocalizeAgent.
The full method, evaluation, and results are in the paper (Batole et al., 2025).