Zimu Wang

Zimu Wang

UC Berkeley CS undergraduate building self-improving and reliable AI agents.

Zimu Wang

I am a computer science undergraduate at UC Berkeley, graduating in May 2027. I recently completed an LLM engineering internship at ByteDance, where I led development of RSIHub and worked on practical agent systems for self-improvement, long-horizon tasks, and code-oriented evaluation. My research focuses on self-improving agents, code intelligence, and reliable AI systems.

Resume PDF | Google Scholar | GitHub | LinkedIn

Research Interests

  • Agent self-evolution and automated improvement
  • Code intelligence and software-engineering agents
  • AI-agent evaluation
  • Reinforcement learning and post-training
  • Reliable long-horizon agents

Selected Work

  • RSIHub GitHub stars — Lead developer; drove the project from 0→1. An open-source recursive self-improvement framework that evolves agent prompts, skills, harnesses, and code while preserving frozen evaluators, declared mutation boundaries, and auditable Git-based experiment lineage. The accompanying paper was accepted to the NeurIPS 2026 Meta-Agents Workshop.
  • EffiSkill: Agent Skill Based Automated Code Efficiency Optimization — Under review via ACL Rolling Review. First-author work on automated code-efficiency optimization that mines reusable Operator and Meta Skills, retrieves relevant skills for new programs, and composes optimization plans without execution feedback.

Selected Experience

  • LLM Engineer Intern, ByteDance — Summer 2026: drove RSIHub’s development as lead developer and built agent skills, composable interfaces, and evaluation infrastructure for long-horizon systems.
  • Researcher, Shanghai Jiao Tong University — Summer 2025–Spring 2026: built EffiSkill and CodeJudge, including scalable code-optimization evaluation and SFT/reinforcement-learning pipelines for code-correctness assessment.

Service

  • Reviewer: NeurIPS 2026 Workshop on Responsible Use of Meta-Agents.
  • Data 8 Tutor, UC Berkeley — Fall 2024: led weekly tutoring sessions, a discussion section, and office hours covering Python-based analysis and visualization, hypothesis testing, regression, and classification.