We are looking for talented individuals to join our team in 2027. As an intern, you will get opportunities to pursue bold ideas, tackle complex challenges, and unlock limitless growth.
Launch your career where inspiration is infinite at our Company.
Successful candidates must be able to commit to an onboarding during the summer 2027. Please state your availability clearly in your resume.
About the team
Our Trust and Safety team is fast growing and responsible for building machine learning models and systems to protect our users from the impact of negative content.
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 moderate the tremendous amount of data generated on the platform.
With our team's continuous efforts, TikTok can provide the best user experience and bring joy to everyone in the world.
Project Overview, Challenges & Value
With the rapid development of
AIGC and the globalization of content ecosystems, content moderation faces three major challenges: evolving policies, surging complexity in multilingual and multimodal content, and upgraded generative adversarial attacks.
The traditional "perception → classification" paradigm has reached its limit.
This topic focuses on two frontier directions: (1) Multimodal moderation foundation model: We study large-scale MoE architecture training and routing optimization, cross-modal alignment and reasoning for multimodality (text/image/video/audio), Unified Understanding & Generation, and high-quality synthetic data generation for moderation scenarios (self-play / adversarial augmentation).
(2) Agentic moderation system: Drawing on advanced agent learning paradigms, it uses reinforcement learning to enhance the agent’s multi-step decision-making capabilities.
It dynamically builds moderation context and integrates a flexible tool ecosystem, enabling autonomous planning, tool collaboration, and interpretable closed-loop reasoning.
This drives a paradigm shift from passive classification to proactive intelligent decision-making in moderation.
Key challenges include:
1. MoE-based multimodal safety foundation model: training stability and routing optimization for large-scale sparse MoE, cross-modal token alignment, and unified architecture design for understanding and generation
2. RL-driven agentic decision-making: end-to-end training of agent multi-step reasoning and tool-call strategies based on GRPO/PPO, overcoming bottlenecks in sample efficiency and training stability
3. Context engineering and tool collaboration: dynamic context assembly, MCP-based tool ecosystem construction, multi-source heterogeneous evidence fusion, and GraphRAG strategy retrieval
4. Generalization and adversarial robustness: generalization across 200+ languages/strategies, adversarial detection of
AIGC content, and design of multi-dimensional reward signals for few-shot scenarios Project Value:
1. Technological leadership: The integration of RL, Agentic, and multimodal foundation models represents the frontier of AI today.
This topic pioneers their application to large-scale content moderation scenarios, with unique advantages in data volume and real-world feedback loops that are impossible to reproduce in pure academic settings.
2. Business value: Serving content safety for billions of users globally and driving the evolution of moderation from dependance on humans and external APIs towards fully automated agentic moderation.
This can directly reduce costs by hundreds of millions of US dollars while improving moderation consistency and response speed.
3. Industry leadership: Mature RL-driven agentic moderation systems do not yet exist in the industry.
This topic could hopefully define the technological paradigm for this direction and produce research outcomes with significant industry influence.