Applied Scientist - Trust and Safety (Multimodal Foundation Model) - Global Frontier Tech Recruitment Program - 2027 Start (PhD)

Written by the T&S Network from TikTok's posting.

Builds machine learning models and agentic systems that moderate harmful content across TikTok's platform, including multimodal (text, image, video, audio) and multilingual content. The team is working on the problem of generative AI content, evolving policies, and adversarial attacks that outpace classic perception-and-classification approaches.

• AI misuse and jailbreaks

What you'd do

• Research and train large-scale sparse MoE multimodal moderation foundation models.
• Design cross-modal alignment and unified understanding and generation architectures for text, image, video and audio.
• Generate high-quality synthetic moderation data through self-play and adversarial augmentation.
• Build reinforcement learning agents that plan multi-step moderation decisions and call tools.
• Develop context assembly and evidence fusion across heterogeneous sources, including GraphRAG retrieval.
• Improve generalization across 200+ languages and detection of AIGC content, including few-shot reward design.

What it asks for

• Doctorate in a computing, data science or AI discipline, finished or close to finishing.
• Strong understanding of current LLM research such as long context, multimodality, alignment and agents.
• Proficiency in Python, Rust or C++.
• Experience with deep learning frameworks such as PyTorch, DeepSpeed, Megatron or vLLM.
• Understanding of distributed computing and performance tuning for training, finetuning and inference.
• Familiarity with PEFT, RL, MoE, CoT or LangChain (preferred).
• Published research or contributions to the LLM community (preferred).

Our best summary of the posting, and it can be wrong. The employer's posting is what counts: read it in full on TikToknorth_east before you apply. If we got something wrong, report this listing.