Zelin Li

Zelin Li

Efficient LLM Training & Inference Systems

  • Integrated Training & Inference

  • Algorithm–System Co-Design

  • LLM Post-Training

  • On-Device AI

CVEmailGitHubLinkedIn
Zelin Li

Education

University of Minnesota, Twin Cities

Bachelor’s in Computer Science

GPA: 4.0/4.0

Beijing University of Posts and Telecommunications

Artificial Intelligence

GPA: 3.85/4.0 · Rank: 1/180

Research Experience

Research Assistant

University of Minnesota · Prof. Caiwen Ding

  • Developed InferZO, a vLLM-based system for zeroth-order fine-tuning with lightweight LoRA-side state and training/inference co-serving.
  • Developing Creeper, a prototype that schedules ready backward computation to reuse GPU weight reads during inference.
  • Studying structured ZO search spaces and mini-batch noise, and exploring LLM execution under memory, bandwidth, and power constraints on edge devices.

Research Assistant

University of Minnesota · Prof. Zirui Liu, with Wenya Xie

  • Implemented and evaluated long-horizon memory systems including HippoRAG and SimpleMem.
  • Contributed to benchmark construction, data auditing, experimental analysis, and manuscript preparation.

Publications & Preprints

  1. Inference-Native Zeroth-Order Optimization for LLMs

    Zelin Li, Caiwen Ding

    Under review at ICLR 2027 · Preprint: arXiv:2605.28760, 2026

  2. Sample-Stable Subspaces and Mini-Batch Noise in Zeroth-Order Fine-Tuning

    Zelin Li, Caiwen Ding

    Preprint, 2026

  3. DynamicMem: A Long-Horizon Memory Benchmark on Evolving Users and Multi-App Behavioral Trajectories

    Wenya Xie, Shengming Zhou, Zelin Li, et al.

    Under review at ACL Rolling Review (ARR), October 2026 · Preprint: arXiv:2606.22877, 2026

Selected Projects

InferZO

An inference-native system for zeroth-order optimization of large language models.

Python · PyTorch · vLLM

Creeper

An ongoing prototype for sharing GPU weight reads between inference and ready backward computation.

Training/inference co-execution · GPU scheduling

HPC Agent Kit

Persistent Slurm GPU workspaces for coding agents, with tmux, fast local scratch, and verified backup/restore.

Python · Bash · Slurm

EasyLocomo

A unified interface and reproducible benchmark framework for long-horizon LLM memory systems.

Python · PyTorch · LLM evaluation

Auto-Chronicle Gaokao

A 120-week state-machine narrative engine with a multi-agent classroom simulation.

Python · LLM APIs · Agents

Honors & Awards

  • National Scholarship, Ministry of Education of China (5/470), 2024
  • ICPC Asia Shenyang Regional Silver Medal (39/300), 2024
  • CCPC Harbin Bronze Medal, 2024
  • Beijing Collegiate Programming Contest Bronze Medal, 2024
  • NOIP Shanxi First Prize, 2021

Technical Skills

Programming
Python, C++
Frameworks
PyTorch, vLLM, Hugging Face Transformers
Tools
Linux, Git