Large Model Application Algorithm Research Scientist - International Content Security

Role Overview

The International Content Security Algorithm Research team at TikTok is committed to fostering a secure and dependable environment for users worldwide. This team focuses on developing and refining machine learning models and information systems that enable early risk detection, rapid incident response, and enhanced threat monitoring capabilities. It spearheads the creation of foundational large-scale models applied across various platform functions such as content moderation, search, and personalized recommendations.

Project Context

Large Language Models (LLMs) have recently made significant strides in natural language processing and AI, showing proficiency in language generation, question answering, and translation tasks. Yet, advancing their reasoning abilities remains a significant challenge. Traditional improvements require substantial supervised fine-tuning (SFT) data, which is costly and hampers scalability. Innovations such as OpenAI's o1 model series have expanded chain-of-thought (CoT) reasoning length to improve results, but scaling such techniques efficiently still remains unresolved.

Alternative approaches including Process-based Reward Models (PRM), Reinforcement Learning (RL), and Monte Carlo Tree Search (MCTS) have been explored, though none match the comprehensive reasoning performance of the o1 series. Emerging research like the DeepSeek R1 indicates that RL may empower LLMs to independently develop robust reasoning capabilities without heavy reliance on expensive SFT data, showing great promise for future advancements.

Key Challenges

• Building effective reward models in RL that accurately represent reasoning quality and dynamically adapt during training.
• Maintaining stable training processes despite limited high-quality SFT data, overcoming the inherent instability and performance risks of RL exploration.
• Extending successful RL methods from structured mathematical and coding tasks to the more complex and open-ended domain of natural language tasks, requiring novel data and algorithmic designs.
• Enhancing reasoning efficiency to balance computational cost with accuracy, exploring techniques such as knowledge distillation and long chain-of-thought strategies for practical deployment.

Candidate Profile

• PhD holder in Computer Science, Electronics, or closely related areas.
• Substantial experience in machine learning domains including computer vision, natural language processing, recommendation systems, or data mining.
• Track record of participation in relevant competitions or industry projects.
• Scholarly publications in top-tier conferences such as KDD, WWW, NeurIPS, ICML, CVPR, ACL, or AAAI.
• Preferred qualifications include research or innovation in large-scale models or reinforcement learning, active open-source contributions, and hands-on deployment experience in production environments.
• Proficiency in programming languages like Python and C++.
• Exceptional analytical and problem-solving abilities paired with a passion for tackling sophisticated challenges.
• Strong communication skills and a collaborative, technology-driven mindset.

About TikTok

TikTok is a premier platform for short-form mobile videos, dedicated to inspiring creativity and delivering joy globally. Headquartered in Los Angeles and Singapore, TikTok operates additional offices worldwide including New York, London, Paris, Berlin, Dubai, Seoul, and Tokyo.

Why Work With Us

At TikTok, we value creativity and innovation, bringing a diverse and global perspective to our work environment. Our teams operate with curiosity, humility, and resilience, constantly iterating to achieve significant breakthroughs. We foster a culture of inclusiveness and growth where every challenge is an opportunity and employees thrive with an 'Always Day 1' spirit.

Diversity and Inclusion

We are passionate about maintaining a workplace that celebrates diversity in all forms and reflects the communities we serve worldwide. TikTok is committed to providing an inclusive space that values unique perspectives and skills.