Semantic-Guided Pruning Approach for Large Language Model-based Class-Level Code Generation

Recent advancements in code generation from natural language have been driven by large language models (LLMs) like GPT-4 and WizardCoder, but current methods often underutilize their capabilities, especially for class-level code generation. The ClassEval benchmark addresses this by focusing on longer, more complex code snippets, but LLMs face challenges with input length constraints. Traditional approaches like zero-shot and few-shot learning struggle with the complexity and length of class-level code, limiting their effectiveness. To overcome this, we propose a semantic-guided pruning method, which optimizes input length through coarse- and fine-grained pruning, retaining only the most semantically relevant tokens. Our method improves LLM performance, increasing pass@1 accuracy by over 4% on the ClassEval benchmark.