ReproCopilot: LLM-Driven Failure Reproduction with Dynamic Refinement

Failure reproduction is a crucial step for debugging software systems, but it is often challenging and time-consuming, especially when the failures depend on complex inputs, states, or environments. In this paper, we present ReproCopilot, a tool that leverages program analysis and a large language model (LLM) to generate failure reproduction code and inputs. ReproCopilot proposes two novel techniques: state-oriented code generation and dynamic refinement that iteratively guide the LLM with program analysis feedback until the generated code can successfully reproduce the failure. We evaluate ReproCopilot on 37 real-world cases from 15 open-source projects, and show that it can reproduce 78% of them, significantly outperforming the-state-of-the-art solutions.