Abstract: In recent years, large language models (LLMs) have demonstrated remarkable capabilities in tasks such as code generation, program repair, and software design. This article reviews the evolution of LLMs in software generation and highlights their potential to improve development efficiency and lower technical barriers. It focuses on three core challenges that remain in complex software generation. In terms of scalability, LLMs still face limitations in output length, logical coherence, and architectural consistency when handling large-scale, modular, and long-lifecycle software engineering projects. In terms of trustworthiness, it is difficult to guarantee the correctness, security, and compliance of generated code, and reliable mechanisms for validating the model’s understanding of requirements are still lacking. In terms of stability, output variability, sensitivity to prompts, and insufficient cross-scenario generalization constrain their deployment in critical systems. The article further explores possible responses to these challenges through hybrid AI approaches, integration with formal verification, and human–machine collaborative design, and offers a perspective on the future role of LLMs in software engineering: they are not a “silver bullet”, but an augmentation tool that must be deeply integrated with traditional software engineering methods.