Interpreting and Steering for Safe and Correct Code Generation

Hao Yan, Ziyu Yao

Large language models (LLMs) frequently generate source code containing vulnerabilities, yet little work studies the internal mechanisms that distinguish safe from vulnerable gener…