The Stanford Center for Research on Foundation Models (CRFM) is an academic research center at Stanford University focused on foundation models. It coined the term "foundation models" in 2021, publishes the annual Foundation Model Transparency Index (FMTI), and has contributed to transparency-centered AI governance through both empirical measurement and policy engagement.
| Parent institution | Stanford HAI (Human-Centered AI Institute), Stanford University (Stanford HAI) | |
| Established | 2021, with the report On the Opportunities and Risks of Foundation Models | |
| Director | Percy Liang | |
| Society Lead | [[rishi-bommasani | Rishi Bommasani]] |
Overview
CRFM is an academic center producing peer-reviewed research on foundation models alongside explicit engagement with governance, a combination that distinguishes it from pure machine-learning labs and from think tanks. Its FMTI methodology and transparency focus are among the mechanisms by which academic research has fed into the transparency-oriented strand of AI policy.
Research and activities
Coining "foundation models" (2021)
CRFM's founding paper, On the Opportunities and Risks of Foundation Models (August 2021, roughly 100 co-authors), introduced the term "foundation models" for what are now called frontier large language models. It was published before ChatGPT (November 2022), establishing the academic vocabulary ahead of the consumer-AI surge.
Foundation Model Transparency Index (FMTI)
The FMTI, published annually since 2023, is an empirical measurement of frontier labs across more than 100 transparency indicators. Scores span ten areas: data (training sources, methodology, data-work labor), labor (compute and human labor used), compute (resources used), methods (training methods disclosed), model access (API, weights, model cards), risks (safety disclosures, red-teaming), mitigations (safeguards, filters), distribution (user access models), usage policy (terms of service), and feedback (user feedback mechanisms).
The FMTI is widely cited in policy arguments for transparency requirements, including in connection with SB 53, the EU AI Act, and the GPAI Code of Practice.
Holistic Evaluation of Language Models (HELM)
HELM is an open-source benchmark suite for evaluating LLMs across multiple capability and safety dimensions. It is a precursor to much of the current benchmarks ecosystem.
California Frontier AI Policy Working Group (2025)
Bommasani's lead-writer role on the working group made CRFM a principal academic contributor to *The California Report on Frontier AI Policy*.
Other work
The Data Provenance Initiative is data-lineage research informing training-data regulation (see AB 2013 — Training Data Documentation (California)). Ecosystem Graphs maps the foundation-model supply chain and its dependencies. CRFM's policy engagement includes repeated testimony in US Senate AI Insight Forums, collaborations with the UK AISI, and EU AI Act consultations.
Relationships
- leads: Rishi Bommasani (Society Lead).
- part-of: Stanford HAI — parent Institute.
- supports: The California Report on Frontier AI Policy — core academic contributor.
- supports: California SB 53 — Transparency in Frontier AI Act / EU AI Act (Regulation 2024/1689) / EU General-Purpose AI Code of Practice (Final Version, 2025) — policy uses of FMTI + framework.
- related: A Framework for AI Development Transparency (Anthropic) — parallel industry-side transparency work.
- related: Frontier Compliance Framework (February 2026) — operationalizes CRFM-style transparency requirements.
- related: AI Benchmarks and Evaluation — HELM and related benchmark work.