Large Model Application Algorithm Research Scientist - International Content Security Algorithms

About the Team and Role

The International Content Security Algorithm Research Team at TikTok is committed to protecting the safety and integrity of our global user base. This team develops and refines machine learning models and information systems designed to detect risks proactively, respond quickly to incidents, and closely monitor potential threats. A key focus is the advancement of foundational large language models (LLMs) utilized across various product features such as content moderation, search, and recommendations. The team's research priorities include addressing challenges related to data compliance, enhancing model reasoning capabilities, and optimizing multilingual performance.

Project Context

Recent strides in large language models have significantly advanced natural language processing applications including language generation, question answering, and translation. Despite progress, reasoning remains an area requiring further development. Prevailing methods typically depend on large volumes of high-quality supervised fine-tuning (SFT) data, which are difficult and costly to obtain, limiting scalability. Emerging techniques like OpenAI's o1 models emphasize extended Chain-of-Thought (CoT) reasoning but scaling these solutions efficiently is an unresolved issue. Research into alternative strategies such as Process-based Reward Models (PRM), Reinforcement Learning (RL), and Monte Carlo Tree Search (MCTS) shows promise but has yet to match the reasoning performance of leading models. Notably, the DeepSeek R1 study illustrates RL's potential to independently cultivate reasoning skills within LLMs without extensive reliance on costly SFT data.

Project Challenges

• Crafting reliable reward models that accurately measure reasoning effectiveness and adapt dynamically during training to refine model performance.
• Ensuring stable RL training despite absence of abundant SFT data, and preventing performance degradation by developing robust training regimens.
• Extending RL reasoning techniques from domains like mathematics and code, which have richer CoT data, to more complex and open-ended natural language tasks by innovating data designs and methodologies.
• Enhancing reasoning efficiency to balance computational cost with accuracy, exploring approaches such as knowledge distillation and extended Long Chain-of-Thought techniques.

Candidate Qualifications

• PhD in Computer Science, Electronics, or related fields.
• Comprehensive experience in machine learning, computer vision, natural language processing, or recommendation systems.
• Proven track record through participation in ML-related competitions or industry projects spanning data mining, CV, NLP, or multimodal applications.
• Publications at respected conferences like KDD, WWW, NeurIPS, ICML, CVPR, ACL, or AAAI.
• Additional assets include research background or innovations in large models or reinforcement learning, contributions to open-source large model projects, and experience deploying such models in production environments.
• Proficient programming skills in Python, C++, or relevant languages.
• Excellent analytical and problem-solving abilities, with a passion for solving complex technological challenges.
• Strong communication skills and a collaborative approach to teamwork.

About TikTok

TikTok is a global leader in short-form mobile video content, inspiring creativity and bringing joy to users worldwide. With headquarters in Los Angeles and Singapore and offices across major cities globally, TikTok fosters a diverse and innovative work environment. Our culture values curiosity, humility, resilience, and a shared commitment to continuous learning and impactful innovation. We aim to cultivate a workplace that embraces diversity and inclusion, valuing the unique perspectives and skills of every employee as we strive to make a positive difference for our community and users.