Structure-Guided and Semantics-Enhanced Collaborative Code Generation

To improve the quality of pseudocode-to-C++ code generation, this paper proposes a Structure-guided and Semantics-enhanced Collaborative Generation method (SSCGen). By integrating a lightweight translation model and a large language model, SSCGen enhances both the structural diversity and semantic correctness of generated programs. Specifically, it first uses a Mixture-of-Experts (MoE) module within an encoderdecoder architecture to perform line-by-line translation and generate well-formed code fragments. These fragments are then assembled into complete programs via best-first search. The resulting candidates serve as structural guidance for a large model, which further performs semantics-enhanced generation. Experiments on the SPoC dataset demonstrate that SSCGen significantly outperforms existing methods, achieving Pass@100 scores of 77.7% and 85.9% on two test subsets.