1 paper · 1 filter
Zijie Zhao, Chenyuan Yang, Weidong Wang +3
While recent LLM-based agents can identify many candidate bugs in source code, their reports remain static hypotheses that require manual validation, limiting the practicality of a…