Using large language models (LLMs) to test compilers is promising but faces two major challenges: the generated programs are often too simple, and large-scale testing is costly. We present a new framework that splits the process into an offline phase—where LLMs generate small, diverse code pieces—and an online phase that assembles them into complex test programs.

Our tool, LegoFuzz, applies this method to test C compilers and has discovered 66 bugs in GCC and LLVM, including many serious miscompilation bugs that prior tools missed. This efficient design shows strong potential for broader AI-assisted software testing.

I am an undergraduate student of Software Engineering in Nanjing University. My research interests lie in the Programming Language and Software Engineering. Recently, I have been working on the Rust Programming Language.

I’ll be joining Chinese University of Hong Kong (CUHK) as a Ph.D. student in 2025 Fall, under the supervision of Prof. Shaohua Li.