Summer Yue is an artificial intelligence researcher specializing in AI safety, alignment, and large-scale machine learning systems. She is a Director at Meta's Superintelligence Labs and previously served as the VP of Research and Director of Safety and Standards at Scale AI. ^{[1] [2]}

Education

Yue attended the University of Pennsylvania, where she was enrolled in the Jerome Fisher Program in Management and Technology. She graduated with a Bachelor of Science degree, earning dual concentrations in Computer Science from the School of Engineering and Applied Science and in Economics from The Wharton School. ^{[3] [2] [6]}

Career

Yue began her career with several internships, including roles as a Web Programmer at Oliden Technology, LLC, a Software Engineering Intern at China Petroleum, a Backend Engineer for Microsoft Office, and a Software Engineering Intern in Compliance Engineering at Square. She then joined Google, where she initially worked as a Software Engineer for YouTube's Trust and Safety division, focusing on issues such as misinformation, spam, and hate speech. She later transitioned to Google's AI research divisions, working as a Senior Research Engineer at Google Brain and subsequently as a Staff Research Engineer at Google DeepMind following their merger. During her tenure at Google from 2018 to 2023, she contributed to research on large-scale deep learning models and infrastructure for projects including Gemini, LaMDA, and AlphaChip.

In November 2023, Yue joined Scale AI as the VP of Research and Director of Safety and Standards. In this capacity, she also served as the AI Chief of Staff to founder and CEO Alexandr Wang. She was hired to establish and lead the company's Safety, Evaluations, and Analysis Lab (SEAL), which focused on research into AI model evaluation, red teaming, and scalable oversight. Her work involved managing several Generative AI machine learning teams dedicated to auto evaluations, synthetic data, and post-training data research.

In July 2025, Yue announced her departure from Scale AI to join Meta's newly formed Superintelligence Labs as a Director. Her work at Meta focuses on AI safety and alignment, continuing her research in building trustworthy and reliable AI systems. ^{[1] [2] [3] [4] [5] [6]}

Major Works and Research

During her time at Scale AI, Yue's work centered on establishing robust methods for evaluating and ensuring the safety of large language models (LLMs). Her primary interests include reinforcement learning, interpretability, value learning, adversarial examples, and fairness in large-scale machine learning systems. ^[3]

Safety, Evaluations, and Analysis Lab (SEAL)

As the head of Scale AI's SEAL, Yue led initiatives to address research challenges in AI safety. A key project under her leadership was the creation of the SEAL Leaderboard, a ranking system for LLMs. The leaderboard was designed to use private, expert-vetted datasets that could not be easily "gamed" by model developers training on public benchmarks. It evaluated models on criteria such as instruction following and their propensity to generate harmful responses to specific prompts. The lab also conducted research into the vulnerabilities of AI agents, finding that safety mechanisms in LLMs did not generalize effectively to downstream browser agents. ^{[5] [1]}

WMDP Safety Benchmark

Yue drove a partnership between Scale AI and the Center for AI Safety (CAIS) to develop the Weapons of Mass Destruction Proxy (WMDP) safety benchmark. This benchmark was created to assess the risks of frontier AI models being misused for malicious purposes, providing a standardized method for evaluating and mitigating potential dangers associated with advanced AI capabilities. ^{[5] [4]}

Selected Publications

Yue has co-authored numerous research papers on AI safety, evaluation, and code generation. Her publications have been featured at conferences such as ICLR and NeurIPS. Notable works include:

• The WMDP Benchmark: Measuring and Reducing Malicious Use with Unlearning (2024)
• A Careful Examination of Large Language Model Performance on Grade School Arithmetic (2024)
• LLM Defenses Are Not Robust to Multi-Turn Human Jailbreaks Yet (2024)
• Aligned LLMs Are Not Aligned Browser Agents (2025)
• Planning in Natural Language Improves LLM Search for Code Generation (2025)
• Humanity's Last Exam (2025)

This list represents a selection of her published research. ^[4]

Public Appearances

Yue is a scheduled speaker at the SXSW 2025 conference, where she will present a talk titled "Beyond the Hype: Building Reliable and Trustworthy AI." ^[5]