Jingjing Xie 谢晶晶

First-Year PhD Student · CUHK-MMLab

Email Scholar Zhihu WeChat

"And don't worry about losing. If it is right, it happens — the main thing is not to hurry. Nothing good gets away."

— John Ernst Steinbeck
Jingjing Xie
Two principles I believe in
  • The speed and correctness of experiments determine how fast both people and models evolve.
  • Research means focusing on the most bottlenecking problem — bold hypotheses, humbly taking others' advice, analyzing from every angle, and iterating fast.

About

I am a first-year PhD student at the Chinese University of Hong Kong, supervised by Prof. Xiangyu Yue at MMLab@CUHK. Before CUHK, I received my master's degree from MAC@XMU, led by Prof. Rongrong Ji.

My research began in model compression and acceleration, where I led several papers. In 2024, I joined the InternVL group to work on the computer-use capabilities of multimodal foundation models and the scaling properties of computer-use agents — sharpening my understanding of data scaling and why it matters. In 2025, I joined StepFun to explore more general reasoning and agentic capabilities beyond computer-use — such as game playing, tool use, and long-horizon coding. Contributing across the various stages of model training and RL infrastructure there deepened my understanding of the foundation-model system as a whole.

I truly enjoy meeting new people and exploring unexpected paths. If my work speaks to you, or if you see a way we could create something together, drop me a line — I'd genuinely love to hear from you.

News

Some Works

Foundation Models
Model Blog
2026
Step 3.7 Flash: A High-Efficiency Flash Model for Real-World Agents[Blog]
Step 3.7 Group
A high-efficiency multimodal agent model (196B total / 11B active, plus a 1.8B ViT) built for real-world agentic tasks — coding, tool use, web / visual search, and GUI operation.
Tech Report
2025
Step 3.5 Flash: Open Frontier-Level Intelligence with 11B Active Parameters[GitHub]
Step 3.5 Group
A frontier-level open model delivering strong intelligence with only 11B active parameters.
Tech Report
2025
Step3-VL-10B Technical Report[GitHub]
Step3-VL Group
A 10B vision-language foundation model from the Step series.
Tech Report
2025
InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency[GitHub]
InternVL Group
An open-source multimodal foundation model advancing versatility, reasoning, and efficiency.
Agents
arXiv 2025
Thinking by Doing: Building Efficient World Model Reasoning in LLMs via Multi-turn Interaction[arXiv]
Jingjing Xie*, Yan Cai*, Jianjian Sun, Chunrui Han, En Yu, Liang Zhao, Jingcheng Hu, Yinmin Zhang, Haoran Lv, Yuang Peng, Zheng Ge, Xiangyu Zhang, Daxin Jiang, Xiangyu Yue
Introduces WMAct, internalizing a world model in LLMs via efficient multi-turn interaction — using reward rescaling and interaction-frequency annealing so multi-turn tasks can be solved in a single turn.
ICLR 2026
Oral
ScaleCUA: Scaling Open-Source Computer Use Agents with Cross-Platform Data[Code]
Zhaoyang Liu*, Jingjing Xie*, et al.
The first large-scale cross-platform dataset spanning 6 operating systems and 3 GUI-centric task domains, plus a general-purpose agent that completes GUI tasks across environments.
CVPR 2026
OS-Oracle: A Comprehensive Framework for Cross-Platform GUI Critic Models[arXiv]
Zhenyu Wu*, Jingjing Xie*, Zehao Li, Bowen Yang, Qiushi Sun, Zhaoyang Liu, Zhoumianze Liu, Yu Qiao, Xiangyu Yue, Zun Wang, Zichen Ding
A scalable pipeline for synthesizing GUI critic data, a benchmark across Mobile / Web / Desktop, and a two-stage SFT + consistency-preserving GRPO training paradigm.
CVPR 2026
MMBench-GUI: Hierarchical Multi-Platform Evaluation Framework for GUI Agents[Code]
Xuehui Wang*, Zhenyu Wu*, Jingjing Xie*, et al.
A hierarchical framework and multi-platform dataset for systematically assessing GUI agents' capabilities across diverse operating systems.
Tech Report
2025
Step-GUI Technical Report[arXiv][GitHub]
Step-GUI Group
A family of 4B / 8B multimodal models for GUI automation, trained with a self-evolving pipeline and a Calibrated Step Reward System — alongside GUI-MCP and the AndroidDaily real-world mobile benchmark.
Model Compression & Acceleration
ICML 2025
GS-Bias: Global-Spatial Bias Learner for Single-Image Test-Time Adaptation of Vision-Language Models[Code]
Zhaohong Huang, Yuxin Zhang, Jingjing Xie, Fei Chao, Rongrong Ji
Unifies global and regional semantics and optimizes at the logit level, enabling lightweight test-time adaptation with reduced inference overhead.
AAAI 2025
TextRefiner: Internal Visual Feature as Efficient Refiner for Vision-Language Models Prompt Tuning[Code]
Jingjing Xie, Yuxin Zhang, Jun Peng, Zhaohong Huang, Liujuan Cao
Equips text prompts with fine-grained visual concepts via a local cache, feature aggregation, and feature alignment — adding local grounding that global prompt tuning lacks.
ACM MM 2024
Advancing Multimodal LLMs with Quantization-Aware Scale Learning for Efficient Adaptation[Code]
Jingjing Xie, Yuxin Zhang, Mingbao Lin, Liujuan Cao, Rongrong Ji
Uses quantization-aware scale learning to fine-tune quantized MLLMs, matching full-precision LLaVA-13B while cutting VL tuning time and GPU cost up to 1.4×.
CVPR 2024
UniPTS: A Unified Framework for Proficient Post-Training Sparsity[Code]
Jingjing Xie, Yuxin Zhang, Mingbao Lin, Zhihang Lin, Liujuan Cao, Rongrong Ji
Achieves practical network sparsity from limited data via a global base-decayed KL objective and an evolutionary search — ~1.54× faster training on ResNet-50 / MobileNet-V2 at SOTA.

* denotes equal contribution.

Experience

2025.08 – now
Research Intern · StepFun
Agentic & multimodal capabilities; contributing to Step 3.5 / 3.7 Flash
2024.10 – 2025.07
Research Intern · Shanghai AI Lab
Agentic reasoning models for computer-use scenarios