Zhen Yang    杨珍

I am a postdoctoral researcher in the Department of Computer Science and Technology, Tsinghua Univerisity, collaborated with Prof. Jie Tang. Prior to that, I received my Ph.D. from the Knowledge Engineering Group (KEG), Department of Computer Science and Technology of Tsinghua University, fortunately working with Prof. Jie Tang. I obtained my master and bachelor degree from Tsinghua University and Xidian University, respectively.

My research interests lie in vision-language models, large language models, and graph representation learning. Currently, I focus on multimodal coding agents and agentic reinforcement learning. I aim to build agentic systems that can reason, plan, and generate executable code cross modalities.

If these areas align with your interests, I welcome you to reach out via email. I am always open to discussing potential collaborations and exploring new ideas together.

Email: yang-zhen [at] mail.tsinghua.edu.cn  /  Google Scholar  /  Github

UI2Code^N is a visual language foundation model trained through staged pretraining, fine-tuning, and reinforcement learning to achieve foundational improvements in multimodal coding, which unifies three key capabilities: UI-to-code generation, UI editing, and UI polishing.

WebVIA is the first agentic framework for interactive and verifiable UI-to-Code generation. While prior vision-language models only produce static HTML/CSS layouts, WebVIA enables executable and interactive web interfaces. The framework consists of three modules: WebVIA-Agent, WebVIA-UI2Code, and Validation Module.

MathSE unifies distilled supervision, an Outcome Reward Model (ORM), and reflection-driven data refresh to progressively enhance math reasoning in multimodal LLMs.

WebSeer is a reinforcement learning framework for training intelligent web-based search agents capable of deeper reasoning, longer tool-use chains, and self-reflective correction. Unlike traditional Retrieval-Augmented Generation (RAG) systems, WebSeer integrates self-reflection into every stage of reasoning, enabling agents to backtrack, reformulate queries, and iteratively improve answers in real-world web environments.

Vision-language models (VLMs) have become a key cornerstone of intelligent systems. As real-world AI tasks grow increasingly complex, VLMs urgently need to enhance reasoning capabilities beyond basic multimodal perception — improving accuracy, comprehensiveness, and intelligence — to enable complex problem solving, long-context understanding, and multimodal agents. Through our open-source work, we aim to explore the technological frontier together with the community while empowering more developers to create exciting and innovative applications. This open-source repository contains our GLM-4.6V, GLM-4.5V and GLM-4.1V series models. For performance and details, see Model Overview. For known issues, see Fixed and Remaining Issues.

We aim to construct a fine-tuning dataset MathVL, and develop a series of specialized mathematical MLLMs MathGLM-Vision with various parameter-scale backbones.

We meticulously construct a comprehensive benchmark, named VisScience, which is utilized to assess the multi-modal scientific reasoning across the three disciplines of mathematics, physics, and chemistry.

We explore the history of negative sampling, categorize the strategies used to select negative samples, and examine their practical applications.

We design the quasi-triangular principle and introduce TriSampler to selectively sample more informative negatives within a prescribed constrained region.

We propose a 2 billion-parameter language model MathGLM that can accurately perform multi-digit arithmetic operations with almost 100% accuracy without data leakage.

We present BatchSampler to sample mini-batches of hard-to-distinguish (i.e., hard and true negatives to each other) instances by leveraging the proximity graph and a random walk with restart.

We propose a novel method ViLTA that utilize a cross-distillation method to generate soft labels for enhancing the robustness of model.

We propose a spatiotemporal aggregation method STAM to efficiently incorporate temporal information into neighbor embedding learning.

We design three region principle to select negative candidate and propose RecNS method to sythesize hard negatives.

We present MixGCF that can study negative sampling by leveraging both the user-item graph structure and GNNs’ aggregation process to design the hop mixing technique to synthesize hard negatives.

We develop a theory and quantify that a nice negative sampling distribution is \( p_n(u|v) \propto p_d(u|v)^\alpha \), \( 0 < \alpha < 1 \). Additionly, we propose Markov chain Monte Carlo Negative Sampling (MCNS), an effective and scalable negative sampling strategy for various tasks in graph representation learning.

• Young Scientists Fund of the National Natural Science Foundation of China (国家自然科学基金青年基金项目C类), 2025.
• General Fund of China Postdoctoral Science Foundation (中国博士后科学基金面上资助), 2025.
• China National Postdoctoral Program for Innovative Talents (2025年度博士后创新人才支持计划), 2025.
• Shuimu Tsinghua Scholar for Postdoc Researcher (清华大学水木学者), 2025.
• Beijing Outstanding Graduate, 2024.
• Huawei Scholarship, 2023.
• Excellent Graduate of Beijing, 2019.
• Outstanding Academic Paper Award, Tsinghua University, 2019.
• 129 Scholarship of Tsinghua University, 2018.
• National Scholarship by Ministry of Education of China, 2018.
• National Scholarship by Ministry of Education of China, 2014.

Thanks to Jon Barron