Our Foundations and Intelligence Service R&D team is fast growing and responsible for building state-of-the-art foundation models, such as LLM, VLM and Omni Models. Our mission is to build a bridge for collaboration between foundation models and downstream business scenarios, and use foundation model powered world knowledge to enhance better user experiences across TikTok, including content moderation, search and recommendations, client AI, etc.
We are looking for researchers in LLM, VLM and Omni Model domain who are experienced in single/multi-modality LLM/VLM pretraining and applications, including evaluations, data processing and recipes for pre-training and post-training, reinforcement learning based alignment, efficient training and inference. There are no doubt a lot of unsolved problems in the LLM domain which could have a huge impact on industry and academia. In TikTok, we have real applications, resources and patience for technology incubation.

Your main responsibilities will be

• Lead, collaborate and execute on research that pushes forward state-of-the-art in multimodal reasoning and generation research
• Enhance VLM with specialized features like OCR and captioning to optimize performance for TikTok business applications
• Explore the model architecture and inference-efficient model design to enable scaled application in downstream TikTok business
• Work closely with cross-functional teams to plan and implement projects harnessing VLMs for diverse purposes and vertical domains
• Extend the insights and impact from industry to academia

Qualifications

Minimum Qualifications
- Ph.D in Computer Science, Data Science, Artificial Intelligence, or a related field
- Strong understanding of cutting-edge LLM research (e.g., long context, multi modality, alignment research, agent ecosystem, etc.) and possess practical expertise in effectively implementing these advanced systems as a plus
- Proficiency in programming languages such as Python, Rust, or C++ and a track record of working with deep learning frameworks (e.g., pytorch, deepspeed, megatron, vllm, etc.).