Large Model Application Algorithm Research Scientist - International Content Security Algorithm Research

Overview

Join the International Content Security Algorithm Research team at ByteDance in Singapore, dedicated to enhancing user safety across international products through advanced machine learning models. This team focuses on building secure, compliant, and high-performance large models supporting content moderation, search, and recommendation tasks.

Research Focus

The role centers on advancing the reasoning capabilities of Large Language Models (LLMs). Despite progress in natural language processing, improving model reasoning remains critical. Current methodologies rely heavily on costly Supervised Fine-Tuning (SFT) data. This project explores reinforcement learning (RL) and novel techniques like Process-based Reward Models and Monte Carlo Tree Search to foster autonomous model reasoning without extensive SFT data.

Challenges

• Develop sophisticated reward models that dynamically adapt to evolving model performances and effectively guide reasoning improvements.
• Ensure training stability in reinforcement learning despite limited high-quality SFT data to avoid fluctuations and performance degradation.
• Extend RL-based reasoning from structured math and coding tasks to complex natural language understanding requiring innovative data and training methodologies.
• Optimize reasoning efficiency by balancing accuracy and computational cost through approaches such as knowledge distillation and Long Chain-of-Thought methods.

Candidate Qualifications

• PhD in Computer Science, Electronics, or a related discipline.
• Significant experience in machine learning, computer vision, natural language processing, or recommendation systems, supported by competition participation or relevant industry projects.
• Published research in top-tier AI, machine learning, or large model conferences (e.g., KDD, WWW, NIPS, ICML, CVPR, ACL, AAAI).

Preferred Extras

• Experience or innovation in large-scale model research or reinforcement learning.
• Proven contributions to open-source large model projects.
• Hands-on experience deploying large models in real-world environments.
• Advanced programming skills in Python, C++, or comparable languages.
• Strong problem-solving abilities and a drive to tackle challenging technical issues.
• Excellent communication, teamwork skills, and enthusiasm for technology.

Company and Culture

Founded in 2012, ByteDance is a global technology company with diverse products like TikTok, Lemon8, and CapCut. The company is driven by creativity and inclusiveness, fostering an "Always Day 1" mindset to innovate continuously and deliver impact. ByteDance values diversity, creating an inclusive environment that respects varied experiences and perspectives worldwide.