Model enhancement method, apparatus, device, storage medium, and computer program product
By constructing a structured directed acyclic graph and using contrastive learning training, large language models can efficiently capture the deep structured logic of complex programs, solving the problem of insufficient code semantic understanding and generation capabilities, and improving the processing accuracy and efficiency of large code models.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING HONGTENG INTELLIGENT TECH CO LTD
- Filing Date
- 2026-04-24
- Publication Date
- 2026-07-24
AI Technical Summary
When dealing with complex and large programs, large language models are limited by the length of the context window, making it difficult to efficiently capture the deep structural design logic of the program, resulting in insufficient code semantic understanding and generation capabilities.
The abstract syntax tree of the code is extracted by the underlying virtual machine tool, a structured directed acyclic graph is constructed, positive and negative sample pairs are generated, and a large language model is trained by contrastive learning to obtain the structured semantic embedding of the code. This is used as input to enhance the training of the large language model of the code and optimize it to adapt to the architecture of different downstream tasks.
It breaks through the limitations of context windows in processing large and complex programs, improves the processing accuracy and efficiency of large code models, achieves accurate learning of deep control flow and data flow logic of programs, and enhances code understanding and generation capabilities.
Smart Images

Figure CN122450459A_ABST