EPFL
academicEPFL (École polytechnique fédérale de Lausanne) is the second pillar (with ETH Zürich) of the Swiss National AI Institute. This entry deliberately excludes that SNAI work: the Apertus base models and the INCLUDE multilingual benchmark are tracked under Swiss AI, and SNAI co-leads Martin Jaggi and Antoine Bosselut carry their SNAI roles there. What remains here is EPFL's independent, non-SNAI research — medical and multimodal foundation models, LLM safety, optimizers, and LLM science.
The sharpest boundary is Meditron, the flagship: an EPFL-origin (epfLLM, pre-SNAI) open medical-LLM lineage co-developed with Yale and the ICRC, later trained on the CSCS Alps supercomputer — the 2026 MeditronFO generation fine-tunes the Apertus base, so we keep the Meditron fine-tunes here and credit the base model to Swiss AI. Other flagship lines: Amir Zamir's VILAB any-to-any multimodal FMs (4M-21, plus FlexTok and AdEMAMix, both co-published with Apple and tracked under Apple); the TML lab's adaptive-jailbreak safety line (simple adaptive attacks, the past-tense jailbreak); Volkan Cevher's LIONS optimizer work (Scion); and Robert West's dlab (GPT-4 persuasion, Nature Human Behaviour 2025).
Martin Jaggi's MLO lab additionally produces general training science — the optimizer benchmark and FineWeb-HQ data selection — that, while non-SNAI-branded, feeds the Apertus pipeline (cross-referenced to Swiss AI where relevant).
People
- Martin Jaggi Google Scholar — Professor, Machine Learning and Optimization (MLO) lab; Steering Committee, Swiss AI Initiative
- Antoine Bosselut Google Scholar — Professor, EPFL NLP lab; Co-Lead, Swiss AI Initiative
- Caglar Gulcehre Google Scholar — Assistant Professor, CLAIRE lab (RL & alignment) (formerly Google DeepMind)
- Amir Zamir Google Scholar — Professor, VILAB (Visual Intelligence & Learning)
- Alexandre Alahi Google Scholar — Professor, VITA lab (Visual Intelligence for Transportation)