Ziheng Qin

Research

Efficient learning through better data and optimization.

My research centers on making learning and agent systems more efficient, especially through dynamic data pruning, data cleaning and growth, data-model co-evolution, scalable optimization, and harness scaling for long-horizon agents.

Research Interests

I work on practical methods that improve training and agent execution efficiency while preserving final quality. Current interests include data-centric AI, optimization for large-scale training, multimodal learning, agent runtimes, and systems for continuously improving datasets and models.

  • Agents
  • Deep learning efficiency
  • Data
  • Optimizer
  • Large-scale training
  • Multimodal learning
  • Data-model co-evolution
441 Google Scholar citations, checked Aug 2026
6 h-index
3 i10-index

Selected Projects

arXiv 2026

StateM

An agent-native runtime for harness scaling on Terminal-Bench 2.1. StateM organizes long-horizon execution around durable states, phase-local context, checked transitions, recoverable runbooks, and versioned procedural practices.

arXiv · Code

ICML 2025

Info-Coevolution

An efficient framework for data-model co-evolution. The method targets active learning settings where datasets grow over time and models must improve under realistic annotation and compute budgets.

Research collaboration between HPC-AI Lab and TikTok; accepted to ICML 2025.

ICLR 2024 Oral

InfoBatch

A lossless training acceleration framework based on unbiased dynamic data pruning. It achieves practical savings across classification, segmentation, detection, vision pretraining, diffusion, and instruction tuning workloads.

ECCV 2024

Dataset Growth

An efficient online algorithm for data cleaning and selection. It maintains growing datasets with awareness of cleanliness, diversity, and sampling value for single-modal and multimodal training.

ICML 2025

MERIT

A maximum-normalized element-wise ratio optimizer for large-batch language model training, designed to improve stability and enable larger effective batch sizes.

Selected Publications

DD-Ranking: Rethinking the Evaluation of Dataset Distillation

Zekai Li, Xinhao Zhong, ..., Ziheng Qin, ..., Yang You, Kai Wang. NeurIPS 2026 ED Track, Oral.

A benchmark for fair evaluation of dataset distillation, separating the effects of knowledge distillation and data augmentation from the quality of distilled data.

arXiv · Code

StateM: Reaching 95.3% Raw Accuracy, or a $15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling

Ziheng Qin, Yaxin Lu, Zhangyang Atlas Wang, Kai Wang. arXiv:2608.15089, 2026.

arXiv · Code

InfoBatch: Lossless Training Speed Up by Unbiased Dynamic Data Pruning

Ziheng Qin, Kai Wang, Zangwei Zheng, Jianyang Gu, Xiangyu Peng, Zhaopan Xu, Daquan Zhou, Lei Shang, Baigui Sun, Xuansong Xie, Yang You. ICLR 2024 Oral.

Scholar entry

Info-Coevolution: An Efficient Framework for Data Model Coevolution

Ziheng Qin, Hailun Xu, Wei Chee Yew, Qi Jia, Yang Luo, Kanchan Sarkar, Danhui Guan, Kai Wang, Yang You. ICML 2025.

Scholar entry

Dataset Growth

Ziheng Qin, Zhaopan Xu, Yukun Zhou, Zangwei Zheng, Zebang Cheng, Hao Tang, Lei Shang, Baigui Sun, Xiaojiang Peng, Radu Timofte, Hongxun Yao, Kai Wang, Yang You. ECCV 2024.

Scholar entry

MERIT: Maximum-normalized Element-wise Ratio for Language Model Large-batch Training

Yang Luo, Zangwei Zheng, Ziheng Qin, Zirui Zhu, Yong Liu, Yang You. ICML 2025.

Scholar entry

Preventing Zero-shot Transfer Degradation in Continual Learning of Vision-language Models

Zangwei Zheng, Mingyuan Ma, Kai Wang, Ziheng Qin, Xiangyu Yue, Yang You. ICCV 2023.

Scholar entry

CowClip: Reducing CTR Prediction Model Training Time from 12 Hours to 10 Minutes on 1 GPU

Zangwei Zheng, Pengtai Xu, Xuan Zou, Da Tang, Zhen Li, Chenguang Xi, Peng Wu, Leqi Zou, Yijie Zhu, Ming Chen, Xiangzhuo Ding, Fuzhao Xue, Ziheng Qin, Youlong Cheng, Yang You. AAAI 2023, Distinguished Paper Award.

Scholar entry

HY-WU (Part I): An Extensible Functional Neural Memory Framework and an Instantiation in Text-Guided Image Editing

Tencent HY Team. arXiv preprint, 2026.

Scholar entry

Research Collaborations

2024 - 2025

TikTok · Data-Model Co-evolution

Proposed and led a collaboration between HPC-AI Lab and TikTok on efficient active learning and data-model co-evolution, with research funding granted by TikTok and a resulting ICML 2025 paper.

2023 - 2024

Ali Tongyi Lab · Data-Centric AI

Worked on data curation and efficient training research, leading to first-author publications at ICLR 2024 and ECCV 2024.

2026

Tencent Hunyuan Lab · Functional Neural Memory

Collaborated on inference-time adaptive neural networks and functional neural memory for text-guided image editing.