Our Trust and Safety RD team is fast-growing and responsible for building machine learning models and systems to identify and defend internet abuse and fraud on our platform. Our mission is to protect billions of users and publishers across the globe every day. We embrace state-of-the-art machine learning technologies and scale them to detect and improve the tremendous amount of data generated on the platform. With the continuous efforts of our team, TikTok can provide the best user experience and bring joy to everyone in the world.

We are looking for people like you with solid experience in designing and deploying state-of-the-art models in the combination of NLP and CV-related areas. This position will work with a team of excellent research scientists and machine learning engineers who can take initiative, design and develop advanced machine learning solutions, and deploy them directly to TikTok's global platform.

Responsibilities - What You'II Do

1. Lead the design, training, and deployment of foundation models, including LLM/VLM, to support a broad range of content safety tasks across modalities. Build general-purpose foundation models with centralized compute and scalable architecture, aimed at improving risk detection, compliance understanding, and moderation automation.
2. Tackle challenges in multilingual, multimodal, and low-resource scenarios by enhancing models’ zero-shot and few-shot generalization across diverse safety domains.
3. Design and maintain a multimodal safety annotation framework, supporting high-quality training data and evaluation signals for both image and video understanding tasks. Explore and implement RLHF strategies to fine-tune model alignment with evolving safety policies and user intent.
4. Collaborate closely with infra, data, and platform teams to optimize large-scale training pipelines, improve model serving efficiency, and leverage centralized GPU resources effectively.
5. Partner with product, policy, and recommendation teams to integrate safety-oriented models into real-world moderation workflows, and continuously optimize for performance and interpretability.