Yingwei Ma

The reproduction of faulty code is a critical step in software debugging, as it enables both the localization of underlying issues and the validation of candidate patches. However, existing approaches often rely on static heuristics or one-shot generation, which limits their ability to handle repository-specific characteristics and the evolving nature of real-world software. To address this challenge, we propose EvoReproducer, a multi-agent online learning framework that continuously refines its reproduction strategies through reflection on past attempts. Central to EvoReproducer is a hierarchical experience pool that adaptively distinguishes between generalizable knowledge and repository-specific experiences, ensuring effective knowledge reuse while preventing uncontrolled experience growth. This design allows the framework to incrementally accumulate useful experiences within the same repository and dynamically adapt to new issues over time. Experimental results on SWEbench Lite show that EvoReproducer substantially improves the success rate of faulty-code reproduction. Moreover, when integrated with downstream debugging agents on SWE-bench Lite, the reproduced tests lead to notable improvements in issue resolution, demonstrating the practical benefits of online experience accumulation for software engineering workflows.

Large language models (LLMs) have demonstrated remarkable performance in code generation, significantly enhancing the coding efficiency of developers. Recent advancements in LLM-based agents have led to significant progress in end-to-end automatic software engineering (ASE), particularly in software maintenance (e.g., fixing software issues) and evolution (e.g., adding new features). Despite these encouraging advances, current research faces two major challenges. First, state-of-the-art performance primarily depends on closed-source models like GPT-4, which significantly limits the technology’s accessibility, and potential for customization in diverse software engineering tasks. This dependence also raises concerns about data privacy, particularly when handling sensitive codebases. Second, these models are predominantly trained on static code data, lacking a deep understanding of the dynamic interactions, iterative problem-solving processes, and evolutionary characteristics inherent in software development. Consequently, they may face challenges in navigating complex project structures and generating contextually relevant solutions, which can affect their practical utility in real-world scenarios. To address these challenges, our study adopts a software engineering perspective. We recognize that real-world software maintenance and evolution processes encompass not only static code data but also developers’ thought processes, utilization of external tools, and the interaction between different functional personnel. Our objective is to develop an open-source large language model specifically optimized for software improvement, aiming to match the performance of closed-source alternatives while offering greater accessibility and customization potential. Consequently, we introduce the Lingma SWE-GPT series, comprising Lingma SWE-GPT 7B and Lingma SWE-GPT 72B. By learning from and simulating real-world code submission activities, Lingma SWE-GPT systematically incorporates the dynamic interactions and iterative problem-solving inherent in software development process—such as repository understanding, fault localization, and patch generation—thereby achieving a more comprehensive understanding of software improvement processes. We conducted experimental evaluations using SWE-bench-Verified benchmark (comprising 500 real GitHub issues), recently proposed by OpenAI. The results demonstrate that Lingma SWE-GPT 72B successfully resolves 30.20% of the GitHub issues , marking a significant improvement in automatic issue resolution (22.76% relative improvement compared to Llama 3.1 405B), approaching the performance of closed-source models (31.80% issues of GPT-4o resolved). Notably, Lingma SWE-GPT 7B resolves 18.20% of the issues, surpassing the 17.20% resolution rate of Llama 3.1 70B, highlighting the potential for applying smaller models to ASE tasks.