Harvard
academicHarvard is comparatively thin on frontier language models but exceptionally strong on scientific foundation models, concentrated in two hubs spanning Harvard Medical School and its hospitals. The Mahmood Lab (HMS / Brigham & Women's / Broad) runs the field's dominant computational-pathology FM program — TITAN, THREADS, KRONOS, and the widely-adopted UNI/CONCH encoders (hundreds of thousands of HuggingFace downloads combined). The Zitnik Lab (HMS-DBMI, mims-harvard) builds therapeutic and molecular FMs and agents: TxAgent + ToolUniverse, ATOMICA, and ProCyon.
The Kempner Institute (co-directed by Sham Kakade, with Cengiz Pehlevan and Boaz Barak) supplies the ML-theory, scaling-law, and training-infrastructure work — Loss-to-Loss Prediction, Archetypal SAE interpretability, and the KempnerForge training stack. Supporting infrastructure like the HEST / TRIDENT pathology toolkits underpins benchmarking across the pathology-FM field.
Industry entanglement is lighter than at peer labs but present: Kakade is also an Amazon Scholar and Barak splits time with OpenAI on alignment and safety. The clinical FMs stay firmly Harvard/Mass General Brigham; note that the TxAgent policy model is a fine-tuned Llama-3.1-8B while ToolUniverse itself is Harvard-original.
Your notes
People
- Sham Kakade Google Scholar — Rampell Professor of CS & Statistics; Co-Director, Kempner Institute (formerly Amazon Scholar (ongoing))
- Cengiz Pehlevan Google Scholar — Associate Professor of Applied Math (SEAS); Kempner Institute faculty
- Boaz Barak Google Scholar — Gordon McKay Professor of CS (SEAS); Kempner faculty (formerly OpenAI (partial leave, alignment/safety))
- Faisal Mahmood Google Scholar — Associate Professor, HMS / Brigham & Women's / Broad; leads Mahmood Lab
- Marinka Zitnik Google Scholar — Associate Professor, HMS-DBMI (mims-harvard); Kempner associate faculty