Curriculum Vitae
Ziheng Qin
Researcher in agents, efficient deep learning, data-centric AI, large-scale training, and multimodal learning.
Profile
I work on practical methods for making learning and agent systems more efficient, including unbiased dynamic data pruning, data cleaning and selection, data-model co-evolution, optimization for large-scale training, and harness scaling for long-horizon agents.
- Agents
- Deep learning efficiency
- Data-centric AI
- Large-scale training
- Multimodal learning
Education
National University of Singapore
Ph.D. student in Computer Science.
University of Southern California
Master in Computer Science.
University of Michigan, Ann Arbor
B.S.E. in Computer Science.
Shanghai Jiao Tong University
B.S.E. in Electronic and Computer Engineering.
Industry Experience and Research Collaborations
Meta Platforms, Inc.
Research Scientist. Working on research projects with potential product applications, including modern recommendation architectures and AI agents/workflows for engineering production.
Manager: Chao Yang.
TikTok Pte. Ltd.
Research Engineer Intern. Proposed and led research on efficient data-model co-evolution for continually improving models and maintaining high-quality growing datasets with affordable annotation and cleaning cost.
Manager: Kanchan Sarkar. Resulting paper accepted to ICML 2025.
Ali Tongyi Lab
Research collaboration on data curation for data-centric AI. Proposed ideas, conducted experiments, and wrote research papers.
Instructor: Yang You. Manager: Baigui Sun. First-author papers at ICLR 2024 Oral and ECCV 2024.
Google LLC
Software Engineer Intern. Designed and implemented a configurable Terraform-based system for automatically setting up VM-based distributed training environments on cloud infrastructure.
Manager: Yixin Shi.
Research Experience
StateM
Agent-native runtime for harness scaling on Terminal-Bench 2.1, organizing long-horizon execution around durable states, phase-local context, checked transitions, recoverable runbooks, and versioned procedural practices.
InfoBatch
Lossless training acceleration by unbiased dynamic data pruning, with 20% to 40% savings across multiple training settings and negligible overhead.
Info-Coevolution
Efficient active learning framework for data-model co-evolution, reducing annotation cost while preserving model quality on classification, segmentation, and real-world in-house data.
Dataset Growth
Online data cleaning and selection algorithm for growing datasets, improving data quality and sampling efficiency for single-modal and multimodal tasks.
MERIT
Large-batch language model training optimizer that uses finer-grained trust ratios and enables larger batch sizes without performance degradation.
Skills
- Python
- C/C++
- Java
- Linux/CentOS
- Deep Learning
- Distributed Training
- Data Mining
- Spark
- XGBoost
- MySQL
- LaTeX
- HTML/CSS/JavaScript
Honors, Awards, and Funding
NeurIPS ED Track Oral
DD-Ranking: Rethinking the Evaluation of Dataset Distillation accepted for an oral presentation at NeurIPS 2026 ED Track, one of 15 oral papers out of 3,757 valid submissions.
ICLR Oral
InfoBatch accepted as an oral presentation at ICLR 2024.
AAAI Distinguished Paper Award
CowClip received the AAAI 2023 Distinguished Paper Award.
TikTok Research Funding
Granted 200k SGD research funding from TikTok for efficient data-model co-evolution.
SJTU-UMJI Summer Design Expo
Golden Prize for the Cloud Gaming Architecture based on StarlingX and Akraino project.