Kejuan Yang

Research Scientist @ TikTok

About

Hi there! My name is Kejuan Yang. I work as a Research Scientist at TikTok TnS team, where my research focuses on pre-training and fine-tuning multimodal large language models for video understanding and reasoning. I graduated and received a dual-degree master's in Computer Science at the University of Washington, in collaboration with Tsinghua University, where I specialize in Natural Language Processing and Machine Learning. Before that, I obtained my Bachelor's degree in EECS from Beijing Institute of Technology. With a keen interest in agent-driven AI, I have worked as MLE @Tiktok, RA @ UW NLP. Feel free to connect with me for any collaborative opportunities or just to exchange insights about the world of AI!

Country

United States

City

San Jose

Industry

Computer Software

Skill

Probabilistic Models, Data Structures, Data Analysis, Machine Learning, Deep leaning, Generative models, Deep Learning, Robot Manipulation, Computer Vision, Visual language model, Multimodality, C (Programming Language), C++, Amplify CLI, Node.js, Next.js, Django, Amazon Web Services (AWS), React.js, AngularJS

Experience

Research Scientist

TikTok

2025-4 - Present · 1 yr 6 mos

San Jose, California, United States

Post-training and evaluating Foundational Multimodal Large Language Models(MLLMs) for fine-grained video captioning

Machine Learning Engineer

Microsoft

2024-9 - 2025-3 · 7 mos

Seattle, Washington, United States

• Implemented Small Language Models as action planner in robot tabletop manipulation tasks through few-shot learning. • Constructed a 3D computer vision dataset to benchmark various open-source SLMs like Phi-3 and Llama-3.1 on embodied reasoning tasks using chain-of-thoughts.

Machine Learning Engineer Intern

TikTok

2024-6 - 2024-11 · 6 mos

San Jose, California, United States

• Quantified the impact of input frame length on long video understanding for Vision Language Models. • Designed and developed a GPU-efficient, tuning-free CLIP-based keyframe selection method that reduced memory usage by 75% while improving accuracy. • Benchmarked the proposed method against state-of-the-art inference-time frame compression methods.

Research Assistant

University of Washington

2024-1 - 2024-6 · 6 mos

Seattle, Washington, United States

• Designed and built a novel benchmark to evaluate LLM-based agents on ML and NLP research repositories. • Expanded benchmark task diversity via data synthesis, scaling from 45 samples to 152 sub-problems and 604 auto-generated examples. • Developed workflow to automate benchmarking execution using proprietary and open-source LLMs.

Research Assistant

Tsinghua University

2023-3 - 2023-8 · 6 mos

Beijing, China

- Researched an innovative CoT application of LLM to complex web-browsing tasks, such as ticketing and online shopping. - Designed and implemented web application to support user instruction in Python. Integrated with LLM for generating a sequential action plan and a Selenium-based web-bot for executing the action plan. - Implemented algorithm workflow and tooling for evaluating LLMs in Python. Integrated the workflow with 25 models such as GPT-4 to automate benchmarking execution. - Released the AgentBench benchmarking for academic use, received 2,500+ stars on GitHub.

Research Assistant

Beijing Institute for General Artificial Intelligence (BIGAI)

2021-10 - 2022-6 · 9 mos

Beijing, China

- Designed and developed a simulation platform to facilitate robotic manipulation and interaction using C# and Unity3D. - Developed rule-based algorithm to enable opening and closing of articulated containers and control of containees. - Identified the issue where robotic arms would collide with container lids during object manipulation. Implemented obstacle-avoidance path planning algorithms to resolve the issue according to the real-time 3D point cloud of the scene. - Integrated VR interaction to enable first-view interaction with objects.