Associate Professor, Shanghai Jiao Tong University
3 papers at NeurIPS 2025
Code Graph Models (CGMs) innovatively integrate both semantic and structural information from code repositories into LLMs, enabling effective repository-level coding tasks without relying on agents or closed-source models.
We propose a flexible realignment framework that enables efficient and controllable realignment of LLMs during both training and inference, addressing challenges in reasoning efficiency and personalized response balancing.