Code intelligence-oriented cross-task shared optimization method and device, and electronic equipment
By combining hierarchical contextual learning with multimodal contrastive learning of code natural language and a shared parameter transfer strategy, the cross-task generalization ability and semantic alignment accuracy of large language models in the field of code intelligence are improved. This solves the problem of weak cross-task transfer ability of models in existing technologies and achieves efficient cross-task tuning and flexible adaptation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- WUHAN UNIV
- Filing Date
- 2026-02-14
- Publication Date
- 2026-06-30
AI Technical Summary
Existing large language models in the field of code intelligence suffer from problems such as weak cross-task generalization ability, weak cross-task transfer generalization ability, and low semantic alignment accuracy between code and natural language. In particular, the model convergence speed is slow and the performance is limited in small sample scenarios.
By co-training hierarchical context learning and code natural language multimodal contrastive learning, combined with a shared parameter transfer strategy, freezing network parameters, and updating only prefix tuning parameters and low-rank matrix parameters, training data is generated and pre-trained to achieve the model's understanding and alignment of code hierarchy and cross-modal semantics.
It significantly improves the model's cross-task generalization ability in the field of code intelligence, reduces the tuning cost of downstream tasks, improves the model's understanding accuracy and semantic alignment ability of cross-level code structures, and achieves efficient cross-task tuning and flexible adaptation.
Smart Images

Figure CN122308837A_ABST