2602.08561

Abstract

Reproducing computational research is often assumed to be as simple as re-running the original code with the provided data. In practice, this rarely works. Missing packages, fragile file paths, version conflicts, or incomplete logic frequently cause published analyses to fail, even when authors share all materials. This study asks whether large language models (LLMs) and AI agents can help automate the practical work of diagnosing and repairing such failures, making computational results easier to reproduce and verify.