Johns Hopkins

Johns Hopkins' Center for Language and Speech Processing (CLSP) and the affiliated Human Language Technology Center of Excellence (HLTCOE) form one of the oldest NLP/speech institutions — Koehn's statistical-MT and Khudanpur's Kaldi lineages, plus the annual JSALT workshops. Its current foundation-model engine is Benjamin Van Durme's group, with PhD student Orion Weller as first author across the line: mmBERT (multilingual ModernBERT-style encoders, 3T tokens / 1,833 languages, ~468K monthly downloads), the Ettin paired encoder/decoder suite (17M–1B), Rank1 — the first test-time-compute reranker — and the instruction-trained retriever Promptriever.

Two more threads are on-thesis: Daniel Khashabi's reasoning and evaluation line (RATIONALYST process-supervision pretraining, Dated Data on effective knowledge cutoffs), and the CCVL medical-vision group (Zongwei Zhou, Alan Yuille) building the AbdomenAtlas dataset + SuPreM pretrained suite and the Touchstone multi-center segmentation benchmark. HLTCOE adds steady CLIR/retrieval infrastructure (ColBERT-X, Rank-K, TREC RAG validators).

Attribution notes: Van Durme's Scholar affiliation is literally "Johns Hopkins University / Microsoft" (confirmed split); DOLOMITES is Google DeepMind's, not JHU's; no frontier-scale CLSP speech artifact shipped in the window — Koehn's MT effort runs largely through Meta collaborations.