With the rapid development of intelligent software engineering and computer science education, automatic programming quality enhancement and quantitative programming performance evaluation have become increasingly critical research directions. Traditional evaluation approaches mainly rely on manual scoring, static code checking tools, and limited test case execution, which are inefficient, subjective, and incapable of capturing deep semantic information and long-range logical dependencies in source code. Meanwhile, existing code optimization methods focus on single tasks such as bug fixing or code summarization, lacking a unified framework that supports both code enhancement and comprehensive performance assessment. To address these limitations, this paper proposes a novel end-to-end deep learning framework named CodeProNet for jointly enhancing programming quality and evaluating programming performance. The model integrates multi-modal feature extraction, semantic-aware graph representation, multi-scale Transformer encoding, and contrastive-learning-based performance prediction. Specifically, we design a semantic-structure fused code representation that combines lexical sequence information, abstract syntax tree (AST) structure, and data-flow graph (DFG) semantics to fully encode intrinsic characteristics of source code. A multi-scale Transformer encoder is introduced to capture both local syntactic patterns and global logical dependencies. Furthermore, a dual-task learning mechanism is constructed to simultaneously optimize code enhancement and performance evaluation. Extensive experiments are conducted on three representative datasets: CodeSearchNet, HumanEval, and a self-built enterprise-level annotated programming dataset (Enterprise Programming Dataset (EPD)). Quantitative results demonstrate that CodeProNet achieves 92.3% accuracy in programming performance grading, 13.7% code error rate, and 85.7% Pass@1 in code functional correctness, significantly outperforming baseline models including CodeBERT, GraphCodeBERT, and CodeT5. Ablation studies verify the effectiveness of each core component. This work provides a unified, scalable, and interpretable solution for intelligent programming education, automated code review, and developer capability evaluation.
Yue Wang (Fri,) studied this question.