Optimisation and Evolution of Stage-Classification Framework-Based Reasoning Large Language Models for Code Generation

Large language models (LLMs) encounter challenges such as logical errors and non-standard syntax in code generation. Although existing research has proposed various code optimisation approaches, there remains a lack of a unified classification framework to systematise their technical lineage and delineate their applicable boundaries. This paper combines the optimisation phases to systematise and summarise research on reasoning-based LLMs in the domain of code generation, categorising various optimisation methods into three stages: pre-generation optimisation, in-generation optimisation, and post-generation optimisation. Pre-generation optimisation, exemplified by the Self-Planning approach, reduces complexity through macro-level decomposition of coding tasks. In-generative optimisation, exemplified by reinforcement-learning-based test feedback CodeRL and natural-language-based reflective feedback Reflexion, enables real-time feedback and remediation. Post-generation optimisation enhances the quality of the final code generation through three core pathways: code filtering, cooperative evolution, and process rewards. The analysis shows that various methods can enhance code generation quality in specific scenarios, but the differences in computational cost, model dependency and applicability are significant. Constrained by testing benchmarks, model scale, and computational costs, in the future, the focus should be on realistic testing benchmarks, lightweight and efficient optimisation pathways, and multi-stage fusion frameworks to drive the high-quality, practical development of code generation.