Zhiheng Lyu
Incoming CS PhD student at UIUC · AI for Software Engineering
Hi, I’m Zhiheng. Welcome to my personal website! This fall I’m joining the University of Illinois Urbana-Champaign (UIUC) as a Computer Science PhD student, advised by Professor Lingming Zhang. My research focuses on AI for Software Engineering, particularly agentic post-training, evaluation, and task synthesis.
I completed my M.Math. in Computer Science at the University of Waterloo (2026), supervised by Professor Wenhu Chen at the TIGER Lab, with a thesis on post-training LLMs as software engineering agents. Before that, I did my undergrad at the University of Hong Kong, where I entered the ICPC World Finals and won two regional gold medals, and worked with research groups at Berkeley, ETH Zürich, and the University of Michigan.
Highlights
- Aug 2026Joining UIUC as a CS PhD student, advised by Prof. Lingming Zhang.
- Jul 2026Released Tencent WorkBuddy Bench — 260 contamination-resistant real-world tasks across Code, Web, Office, and Security.
- Jun 2026VerlTool accepted to TMLR, after winning the Best Paper Award at the SPOT Workshop @ ICLR 2026.
- May 2026MiniMax-M2 tech report released — 69% Pass@1 on SWE-bench Verified, #2 on MultiSWE and TerminalBench.
- May 2026SWE-QA-Pro accepted to ACL 2026 Findings — our trained 8B model surpasses GPT-4o on repository-level QA.
Experience
Tencent (Feb – Jul 2026) — Intern (Qingyun Program) on WorkBuddy, working closely with Ke Li and Chao Peng. I owned the team’s evaluation platform and infrastructure (3× evaluation volume), explored task synthesis from production user data, and co-led the release of Tencent WorkBuddy Bench (code).
MiniMax (May 2025 – Feb 2026) — Research Scientist Intern on the base model team, contributing to code-agent post-training for MiniMax M1 through M2.5. M2 achieved 69% Pass@1 on SWE-bench Verified (#2 on MultiSWE and TerminalBench). I led large-scale SWE data synthesis (36K verifiable tasks from 5K+ sandbox environments), built the rubric-based evaluation benchmark that became the team’s core metric for code-agent user experience, and led CodeMirror ToolScaling for M2.5.
Research
Agentic Post-Training
I work on the full pipeline of post-training for software engineering agents. As a co-first author and core developer of VerlTool (TMLR 2026; Best Paper Award at the SPOT Workshop @ ICLR 2026), I built the stateful environment interaction protocol and SWE agent training pipeline. In my master’s thesis, RLVR training took Qwen3-8B from 10.4% to 19.5% on SWE-bench Verified. I also co-first authored SWE-Next (2,308 verifiable tasks mined from 311 real repositories) and BrowserAgent (TMLR).
Benchmarks & Evaluation
As models get stronger, the definition of tasks becomes increasingly important. SWE-QA-Pro (ACL 2026 Findings) builds a contamination-resistant repository-understanding benchmark from long-tail repositories; WorkBuddy Bench reconstructs real-world tasks from production scenarios; PixelWorld (TMLR) probes reasoning by converting text into pixels.
Earlier, I worked on causal reasoning and knowledge methods for LLMs: FactTrack (NAACL 2025 Oral) for time-aware world state tracking in story outlines, Corr2Cause (ICLR 2024) for testing pure causal inference, and psychologically-inspired causal prompting. I also contributed to StructEval and VideoScore.
For future research directions, see my Research Statements; my complete background is in my CV.
Selected Publications
An open benchmark of 260 contamination-resistant, real-world tasks spanning Code, Web, Office, and Security.
Technical report on the MiniMax-M2 MoE series (229.9B total / 9.8B active), built end-to-end for agentic deployment. 69% Pass@1 on SWE-bench Verified, #2 on MultiSWE and TerminalBench.
An execution-grounded framework for scalable SWE task and trajectory collection: 2,308 verifiable tasks mined from 311 real repositories.
A contamination-resistant benchmark for agentic repository-level code understanding, plus a training recipe that lets an 8B model surpass GPT-4o.
A unified, easy-to-extend tool-agent RL training framework based on verl, supporting agentic RL across code, search, SQL, and SWE environments.
A framework for training web agents that directly interact with browser environments for information-seeking tasks
Converting textual reasoning data into images to probe vision-language model reasoning capabilities
A novel approach to tracking dynamic world states and detecting contradictions in story narratives
This research introduces the first benchmark dataset, Corr2Cause, to test large language models (LLMs) pure causal inference skills.
Contact
Feel free to reach out at zhihenglyu.cs@gmail.com for research collaboration, open-source projects, or mentorship.