New ReCUBE benchmark measures how well AI coding assistants use repository context to generate code

ReCUBE tests LLMs' ability to reconstruct masked files within real-world repositories using only available source files, dependencies, and documentation as context

The benchmark evaluates generated code with usage-aware test cases that verify both internal module logic and cross-file integration, mimicking real software patterns

Researchers introduce the Caller-Centric Exploration (CCE) toolkit, a dependency graph-based tool system designed to enhance LLM code generation capabilities

This is the first benchmark to directly isolate and measure how effectively large language models leverage repository-level context during code generation tasks