Libra Optimizes Agentic LLM Information Retrieval by Training Environment

Agentic Large Language Model (LLM) systems heavily rely on their ability to locate specific information within vast repositories. While synthetic data has been used to train LLMs themselves, less attention has been given to optimizing the agent's operational environment, specifically the repository structure, in a data-driven manner. Libra addresses this gap by presenting a self-evolving framework that integrates mutable "catalogs" into the repository. These catalogs are hierarchical Markdown files that function as navigable indices. Libra employs an LLM-driven optimization loop where a "Prompter" generates synthetic queries, a "frozen Solver" attempts to resolve them by navigating the catalogs, and a "Healer" then rewrites the catalogs in response to any localization failures encountered by the Solver. Evaluations across 12 SWE-bench Lite repositories demonstrated that this environmental "healing" process resulted in continuous, logarithmic improvements in code localization accuracy. Crucially, these environmental enhancements transferred zero-shot to different LLMs and problem sets. The research also showed that a minimalist coding agent equipped with Libra-optimized catalogs outperformed state-of-the-art baselines.