
CS PhD student @UChicago, ex MS @UofIllinois, research intern @IBMResearch homepage: https://nitter.cf/t.co/w1jpH1vEOk
Chicago
Joined October 2021
- Tweets55
- Following256
- Followers171
- Likes684
Pinned Tweet
This post frames verification as a loop: try to prove, fail, use the failure to find the bug, fix, repeat.
When an LLM runs that loop, the key question is how much it learns from each failure. Usually not much. Our two recent papers on why, and what to do 🧵
New on our Frontier Red Team blog: using Claude to formally verify software at large scale.
Formal verification can prove that code is free of certain bugs, but historically it has only been feasible for small codebases (or required a lot of expert time).
With AI models, however, this is beginning to change. In collaboration with @theoremlabs, we used Claude to prove security properties in over one hundred real-world open-source libraries over the span of a few weeks.
Read our research in full here: anthropic.com/research/forma…
4/
TopoVerus (NeurIPS ‘26): a correct counterexample can still be useless. Invariants build on each other, so we train the model to produce counterexamples that actually lead to a proof, establishing foundational invariants first.
1
2
29
5/
As proof generation scales, what the model learns from each failure is what matters most to fix the failure. We found counterexamples are the perfect signals/witnesses to work on for both human and LLMs.
ExVerus: exverus-proof.github.io/
TopoVerus: etha-sun.github.io/papers/To…
w/ @Yuechun_Sun, @yiwu_0810, Rodrigo Caridad, @YuanYongwei, Jianan Yao, Shan Lu, and
@Kexin_Pei
2
38
Excited to share that I’ll be presenting our #ICML2026 paper in Seoul:
ExVerus: Verus Proof Repair via Counterexample Reasoning
📍 Poster: Thu, Jul 9, 5:00–6:45 PM KST, Hall A #2004
💡 DL4C Workshop: Fri, Jul 10, Session 2, 3:20–4:50 PM KST, Hall A (Board 100-119)
Paper: openreview.net/forum?id=FNDk…
Happy to discuss!
This work focuses on LLM-based Verus proof repair. Verus is a practical Rust-based deductive verification framework that has been used to verify large-scale critical systems, such as cluster management controllers (Anvil) and virtual machine security modules (VeriSMo).
In this work, we propose a new paradigm for LLM-based proof repair: repairing a proof via synthesizing, validating, and blocking counterexamples.
1/n
Takeaway: counterexamples are actionable feedback that explains why the proof fails. We make source-level counterexample generation and validation practical by iteratively synthesizing source-level SMT/Z3Py queries and validating them until a good set of counterexamples is found. The counterexamples serve as strong actionable feedback that guides proof repair more effectively.
8/n
I’d be happy to chat about coding agents, LLMs for code, formal verification, Verus/Rust, and counterexample-guided reasoning at ICML!
Huge thanks to my amazing co-authors: @Yuechun_Sun, @yiwu_0810, Rodrigo Caridad, @YuanYongwei , Jianan Yao, Shan Lu, and @Kexin_Pei .
Code: github.com/claudeyj/exverus
#ICML2026 #LLM #FormalMethods #SoftwareEngineering #ProgramVerification #Rust #Verus
9/n
Jun Yang retweeted
I am on the job market - Very happy to give a talk at the UIUC SysNet seminar!
Had an amazing discussion that reinforced my belief: Code generation is maturing; beyond it, the bigger challenge is to manage running systems reliably and efficiently. 🚀
Another poster session on Dec 6 (Sat) 1:20-2:20pm at Exhibit Hall G&H at #NeurIPS2025 . Please drop by!
Drop by our poster #2408 on Dec 3 (Wed) 11– 2 PST at Exhibit Hall C,D,E at #NeurIPS2025 to check our paper "Synthesizing Performance Constraints for Evaluating and Improving Code Efficiency" 🎉📍
We introduce WEDGE: a framework that helps LLMs generate "performance-stressing" tests to expose inefficient code. 🧵👇
4
388
Drop by our poster #2408 on Dec 3 (Wed) 11– 2 PST at Exhibit Hall C,D,E at #NeurIPS2025 to check our paper "Synthesizing Performance Constraints for Evaluating and Improving Code Efficiency" 🎉📍
We introduce WEDGE: a framework that helps LLMs generate "performance-stressing" tests to expose inefficient code. 🧵👇
Paper link: arxiv.org/abs/2505.23471
Repo link: github.com/UChiSeclab/wedge
Benchmark link: github.com/UChiSeclab/perffo…
2
126
Big thanks to my amazing mentor Bogdan @CacheMisses, awesome advisor @Kexin_Pei and collaborators!!! 🙌🎓❤️
#AI #LLM #SoftwareEngineering #CodeGeneration #Fuzzing #Research #NeurIPS2025 #LLM4Code
8/8
1
2
162
We are open-sourcing both the framework and the benchmark! 🎁📂
1️⃣ WEDGE: The source code for synthesizing constraints and guiding fuzzing. 🔗 github.com/UChiSeclab/wedge
2️⃣ PERFFORGE: The resulting dataset of performance-stressing tests. 🔗 github.com/UChiSeclab/perffo…
7/8
1
94



