4th-year PhD student at UIUC. Improve the reliability of software systems by fuzzing, reasoning and verification along with LLMs

Joined June 2021
🚨 We taught LLMs to write bug-finding tools instead of just finding bugs! KNighter synthesizes static analysis checkers from patches & discovered 92 new, long-lived bugs in Linux (77 confirmed, 30 CVEs) 🐛 Paper: arxiv.org/abs/2503.09002 Code: github.com/ise-uiuc/KNighter Presenting TODAY 3:30 PM KST at #SOSP2025!
3
33
3
111
15,706
❓How can we trust AI-generated code? By teaching AIs to write formal proofs! Our paper, AutoVerus, an agentic framework that guides LLMs to generate Verus proofs for Rust code, proving >90% of tasks in VerusBench. 📄 Paper:: dl.acm.org/doi/pdf/10.1145/3… 🏆 Code (Distinguished Artifact): github.com/microsoft/verus-p… 🎤 Talk: Tomorrow (Oct 18) at 4:15 PM SGT: 2025.splashcon.org/details/O… #FormalVerification #RustLang #AI
2
24
2
108
7,404
Big thanks to my amazing collaborators Xuheng Li, @MRHMisu, Jianan Yao, @weidongcui, Yeyun Gong, Chris Hawblitzel, @LahiriShuvendu, Jacob R. Lorch, Shuai Lu, Fan Yang, Ziqiao Zhou, Shan Lu!!!
2
396
🔧 How? Multi-stage synthesis: - Extract bug patterns from patches - Generate state machines & checker logic - Compile to real C++ checkers - Validate & refine automatically The checkers are code, not black boxes—auditable & evolvable!
1
2
1
428
Big thanks to my amazing collaborators @zzjas98, Zichen Xie, Haoyu Li, and @LingmingZhang ! 🎓 Come chat at SOSP if you're interested in LLM × program analysis, or join us for future work! #ProgramAnalysis #LLM #SoftwareEngineering #LinuxKernel
1
1
353
How to leverage the white-box info (i.e. source code) for fuzzing compilers? Check out our work “WhiteFox 🦊: White-Box Compiler Fuzzing Empowered by Large Language Models” at OOPSLA 2024! w/ @yinlin_deng, @lry89757, JIayi Yao, @JiaweiLiu_, @Reyhaneh, and @LingmingZhang (1/N)
1
19
1
66
12,546
We've also made some observations about optimization-related bugs in PyTorch Inductor. We hope these insights will benefit the entire DL compiler community! (8/N)
1
1
2
392
🚨LLMs are Zero-Shot Fuzzers! Excited to share our TitanFuzz🤖 work @issta23: #LLMs can be directly applied for both generative and mutative fuzzing, while being fully automated, generalizable, and applicable to domains challenging for traditional fuzzers (such as DL systems)🧵
2
48
4
202
30,164