CV

Download

Download full resume (PDF)

Education

• University of California, Berkeley - B.A. in Computer Science
• GPA: 3.92 / 4.0
• Expected graduation: May 2027
• Coursework: Neural Networks, Deep Reinforcement Learning, Computer Vision, Optimization, Probability Theory, Databases, Data Structures & Algorithms

Professional Experience

• LLM Engineer Intern, ByteDance - Summer 2026
  • Drove RSIHub’s development as lead developer, delivering an open-source recursive self-improvement framework that evolves agent prompts, skills, harnesses, and code.
  • Architected a modular evolution engine that isolates mutable agent components from fixed evaluators and records every candidate generation in Git, preserving score integrity and commit-level reproducibility.
  • Shipped a plugin operator SDK with automatic discovery, subprocess isolation, and declarative schemas, enabling new operators and strategies without core changes.
  • Demonstrated best full-benchmark gains of +13.5 points on Terminal-Bench 2 with MiniSWE (56.2% → 69.7%) and +9.3 points on Tau3 Banking with Codex (24.7% → 34.0%) across four self-improvement strategies and all 186 tasks in the two benchmark suites.
  • For Simple Long Horizon Agent, shipped an on-demand skill runtime with package discovery, compact prompt menus, explicit skill selection, and a bundled reusable skill library.
  • Unified four agent presets behind a composable AgentSession/toolset API, standardizing session and tool lifecycles while adding config-driven MCP support to evaluation runners.

Research Experience

• Undergraduate Researcher, University of California, Berkeley - Spring 2026; Fall 2026–Present
  • Advised by Professor Avideh Zakhor.
  • Studied whole-body motion-control methods and trained control policies in MuJoCo and Isaac Lab for the Unitree G1 humanoid robot.
  • Investigated simulation-to-real deployment and hardware-integration challenges on the Unitree G1 platform.
• Researcher, Shanghai Jiao Tong University - Summer 2025–Spring 2026
  • Mentored by Yuling Shi under the supervision of Professor Xiaodong Gu.
  • EffiSkill: designed a two-stage LLM framework that mines reusable Operator and Meta Skills from slow/optimized program pairs, retrieves relevant skills, and composes optimization plans for new programs without execution feedback.
  • Built the inference and evaluation stack and benchmarked all 623 EffiBench-X tasks in both Python and C++, with optimization success rates 3.69–12.52 percentage points above the strongest baseline.
  • CodeJudge: built and scaled a pipeline to train Qwen3-8B to assess code correctness with 48K examples on 8×A100 GPUs, using SFT and reinforcement learning, custom rewards, and distributed verl/vLLM rollouts.
  • Raised held-out ground-truth agreement from 54% to 69%, outperforming Qwen3-32B (66%) while using 25% as many parameters.
• Research Assistant, University of California, Berkeley - Fall 2025
  • Advised by Professor Park Sinchaisri.
  • Built end-to-end processing pipelines for a 1.2M-row gig-economy dataset and produced validated, analysis-ready datasets for collaborators.
  • Ran 30+ linear regressions and generated 40+ plots with robustness checks using statsmodels and scikit-learn.

Teaching

• Data 8 Tutor, University of California, Berkeley - Fall 2024
  • Led weekly tutoring sessions, a discussion section, and office hours, teaching Python-based data analysis and visualization alongside hypothesis testing, regression, and classification.

Skills

• Languages: Python, C++, Java, SQL, Bash
• ML / Agents: PyTorch, Hugging Face Transformers, verl, vLLM, SFT, GRPO, RLHF, MCP, agent evaluation
• Infrastructure / Tools: Docker, Weights & Biases, Git, Linux, NumPy, Pandas

Selected Publications

• RSIHub: A Composable Framework for Agent-Evolution Methods Permalink

  Zimu Wang^*, Xiaobo Wang^*, Siyu Ye^*, Jiaqi Li^*, Silin Chen, Yuling Shi, Ruobing Wang, Jingjing Zhang, Chaofan Wang, Kai Wu, Mengnan Qi, Kai Cai, Xiaodong Gu^†, Zilong Zheng^†. NeurIPS 2026 Meta-Agents Workshop. ^* Equal contribution; ^† Corresponding authors.

• EffiSkill: Agent Skill Based Automated Code Efficiency Optimization

  Zimu Wang, Yuling Shi, Mengfan Li, Zijun Liu, Jie M. Zhang, Chengcheng Wan, Xiaodong Gu. Under review via ACL Rolling Review.