Hong Kong University of Science and Technology
Hong KongPh.D. in Computer Science and Engineering
CV
Education
Beijing Institute of Technology
BeijingB.S. in Statistics, School of Mathematics and Statistics
B.Eng. minor in Artificial Intelligence, School of Computer Science
Representative Work
Qwen3.5 / Qwen3.6
Leading Computer Use Agent capability.
Owned RL infrastructure, data quality, CPT/SFT/RL training, evaluation, and internal benchmark loops for the CUA pipeline.
Qwen3-Coder
Contributor, 2025.
Worked on the Browser Use Agent module, browser interaction data construction, capability improvement, training pipeline support, and evaluation.
DataFlow
Core contributor, 2025.
Responsible for the code data pipeline, including code data processing, quality filtering, pipeline orchestration, and preparation of training-ready code data.
Selected Publications
Multimodal GeometryMentorICML 2026
Synthesizing Multimodal Geometry Datasets from Scratch and Enabling Visual Alignment via Plotting Code
Mentored junior collaborator on synthetic multimodal geometry dataset generation and visual alignment with plotting code.
Multimodal LearningFirst authorNeurIPS 2025
Hallucination at a Glance: Controlled Visual Edits and Fine-Grained Multimodal Learning
Introduces the MicroEdit Dataset and a fine-tuning framework that improves MLLM sensitivity to fine-grained visual differences.
Visual ReasoningFirst authorNeurIPS 2025
Multi-Step Visual Reasoning with Visual Tokens Scaling and Verification
Proposes scalable token-based iterative reasoning and verification for complex visual reasoning.
Data SelectionFirst authorACL 2025 Main
Efficient Pretraining Data Selection for Language Models via Multi-Actor Collaboration
Presents a multi-actor collaborative data selection framework for more efficient language model pretraining.
Synthetic DataCo-first authorICLR 2025 Spotlight
LOKI: A Comprehensive Synthetic Data Detection Benchmark using Large Multimodal Models
Benchmarks MLLMs across video, image, 3D, text, and audio synthetic data detection tasks.
Data-Centric AIFirst authorACM Computing Surveys, major revision
A Survey of Multimodal Large Language Model from A Data-centric Perspective
Surveys multimodal large language models from data preparation, evaluation, benchmarks, and future directions.
Internships
- Research Intern
Work on agent capabilities for Qwen models, including Browser Use Agent for Qwen3-Coder and Computer Use Agent for Qwen3.5/Qwen3.6. My work spans data construction, RL infrastructure, model training, evaluation, and failure analysis. - Research Assistant, supervised by Prof. Wentao Zhang
Contribute to DataFlow, focusing on code-data workflows, code data pipeline construction, data processing, and quality filtering for training-ready code data. - Research Intern
Worked on data preparation and selection for LLM pretraining, including data management strategies for InternLM3-8B and Ray-based labeling pipelines for data selection. Led three first/co-first author research projects during the internship. - Research Assistant, supervised by Prof. Bin Cui
Studied transfer learning for Bayesian optimization and hyperparameter tuning. This work led to a KDD 2022 paper on transfer-learning-based search space design, and a first-author survey on transfer learning for Bayesian optimization.
Honors and Awards
- Excellent Graduation Thesis, School of Computer Science, Beijing Institute of Technology, 2023
- Gold Medal, International Genetically Engineered Machine competition, 2021
- Meritorious Winner, Mathematical Contest in Modeling, COMAP, 2021
- Outstanding Student, Beijing Institute of Technology, 2020, 2021, 2022
- First Prize, Chemistry Olympiad, provincial level, 2018