Speculative sampling has been widely applied in natural language processing. By combining the characteristics of code itself, integrating different LLMs can further accelerate code generation.

LLMs have demonstrated proficiency in handling single-task scenarios such as single-file code generation and project-level code completion. However, there is a notable research gap in zero-shot generation of entire projects. The research objective is to construct a benchmark dataset for project-level code generation and explore more efficient methodologies for generating code at the project level.

Different LLMs have their own advantages and disadvantages for different programming problems. How to optimize the selection of LLMs for code generation problems by evaluating the difficulty of the problems and balancing the resource occupation of LLMs is a key point.

LLMs are widely used in code generation, but there is a lack of research on the underlying mechanisms. Does an LLM generate code by leveraging the structure of the code or through semantic information such as variable names?