Stanford University

Center for Research on Foundation Models (CRFM) / Stanford AI Lab

United States · PhD/Postdoc

Stanford CRFM is the core academic center for foundation model research, investigating the safety alignment, evaluation, and governance of large language models. Building upon the deep expertise of the Stanford AI Lab (SAIL), the center produces influential research in RLHF, model evaluation, and safety benchmarking, serving as a critical hub at the intersection of AI safety academia and industry.

学术历史

The Stanford AI Lab (SAIL) was founded in 1962 and is one of the world's earliest AI research institutions. In 2021, John Etchemendy and Fei-Fei Li co-founded CRFM in response to the safety and governance challenges posed by large language models (e.g., GPT-3). The establishment of CRFM marked a systematic academic investment in foundation model safety research, and its HELM evaluation framework has become an important benchmark for model safety assessment. Stanford has a long-standing tradition in reinforcement learning and AI safety, with Emma Brunskill's RL lab and Chelsea Finn's meta-learning research providing foundational techniques for alignment.

当前状态

CRFM currently operates across multiple research directions: foundation model evaluation (HELM), safety alignment, model governance, and multimodal safety. The center shares resources with SAIL and has access to large-scale computing clusters for model training and evaluation. It offers courses such as CS324 (Large Language Models) to cultivate talent in safety research. The center regularly publishes foundation model safety reports and organizes the Stanford AI Safety Workshop.

实验室 / 研究中心

关键人物

标志性成果

• Released the HELM (Holistic Evaluation of Language Models) framework, establishing an academic benchmark for large model safety and capability evaluation
• Proposed the Foundation Models conceptual framework, systematically defining the safety and governance challenges of the large model era
• Pioneering work on the MAML (Model-Agnostic Meta-Learning) algorithm, providing technical foundations for rapid adaptation and value learning
• Theoretical framework for safe reinforcement learning, defining safe exploration boundaries in constrained Markov decision processes

学术资源

证据