Youcheng Sun.

Research

Agents are the new software.

Software used to be written line by line by people. Increasingly, it is written, and run, by AI agents. I study these agents as a new kind of software: how they behave, how to build better ones, and how to be sure the code they produce is correct.

1. 1

   How agents behave

   Agents don't just follow fixed instructions; they plan, adapt, and, when many of them interact, can cooperate or even deceive one another. I study what agents really do, and where that goes wrong.

   Tapas · AAAI 2026  ·  Lying with Truths · ACL 2026 Outstanding Paper
2. 2

   Building better coding agents

   The most useful agents are the ones that write code, and stronger coding agents make stronger general-purpose agents. I work on how they plan their work, reuse learned skills, and how we measure whether they truly help.

   SWE-Manager · Skill Coverage · SWE-Skills-Bench
3. 3

   Verifying agent-written code

   Code shouldn't be trusted just because an agent wrote it. I build tools that automatically check whether generated code is actually correct.

   AProver · Agentic Model Checking · Agentic Planning for Symbolic Execution

PhD Openings

News

• Jul 2026🏆 ACL 2026 Outstanding Paper Award for “Lying with Truths: Open-Channel Multi-Agent Collusion for Belief Manipulation via Generative Montage.”
• Jun 2026Congratulations to Ali Aljaberi! His paper “AIALIB: A Threat Library for AI-Generated Code” is accepted to the ASE 2026 Tools & Datasets Track.
• Jun 2026Our UGRIP 2026 internship team (Dragos-Gabriel Monea, Yilin Liu, Rishika Mehta) won the Best Project Award for Efficient, Adaptive & Extensible Smartphone-Native Agent Runtimes.
• May 2026M.Sc. student Ali Aljaberi graduated and received the Future Tech Titan Award.
• Jan 2026Ph.D. student Hamad Alshehhi received the Outstanding TA Award.

Projects

ClawMobile · TypeScript

An agent-first phone environment for Android: an AI agent becomes the primary interface that observes device state, controls apps with permission, and learns reusable mobile skills, with app-local and Termux/OpenClaw runtimes.

github.com/ClawMobile/ClawMobile ↗

AProver · Python

Agentic prover for AI-generated code: LLM-driven formal verification agents that pair large language models with bounded model checking (CBMC / Kani / JBMC) to auto-generate specifications, classify counterexamples as real or spurious, and report confirmed bugs.

github.com/agentic-prover/aprover ↗

Recent Papers · 2026

1. Lying with Truths: Open-Channel Multi-Agent Collusion for Belief Manipulation via Generative Montage ACL 2026 Outstanding Paper
   J. Hu, X. Huang, Y. Sun, Y. Dong, X. Huang · ACL 2026
2. Agentic Model Checking
   Y. Sun, J. Liu, D. Kroening, J. Xue · arXiv 2605.21434
3. Agentic Planning for Symbolic Execution
   D. Koh, Y. Noller, C. S. Păsăreanu, Y. Sun · arXiv 2608.06397
4. Tapas are Free! Training-Free Adaptation of Programmatic Agents via LLM-Guided Program Synthesis in Dynamic Environments
   J. Hu, Y. Dong, Y. Sun, X. Huang · AAAI 2026
5. ClawMobile: Rethinking Smartphone-Native Agentic Systems
   H. Du, S. Wu, Q. Li, R. Pan, J. Li, Y. Sun, C. J. Xue · EuroMLSys 2026
6. Skill Coverage: A Test Adequacy Metric for Agent Skills
   B. Tan, X. Huang, Y. Sun · arXiv 2606.20659
7. SWE-Skills-Bench: Do Agent Skills Actually Help in Real-World Software Engineering?
   T. Han, Y. Zhang, W. Song, C. Fang, Z. Chen, Y. Sun, L. Hu · arXiv 2603.15401
8. SWE-Manager: Selecting and Synthesizing Golden Proposals Before Coding
   B. Tan, H. Deng, J. Zhang, J. Xu, P. He, Y. Sun · arXiv 2601.22956
9. Specification Vibing for Automated Program Repair
   T. Zhu, L. C. Cordeiro, M. A. Mustafa, Y. Sun · arXiv 2602.08263

Full list of papers on Google Scholar ↗