University of Edinburgh
academicEdinburgh's frontier-AI output centers on the ILCC / EdinburghNLP cluster in the School of Informatics. The flagship line is HPLT — the multi-national High Performance Language Technologies consortium (Edinburgh a key partner with Oslo, CUNI, Turku, and Helsinki) whose open multilingual pretraining corpora scaled from ~7.6T whitespace tokens across 193 languages (v2, ACL 2025) to ~30T tokens across 198 languages (HPLT 3.0, CC0, distilled from 7.2PB of Internet Archive and Common Crawl). Edinburgh is also a partner (Lisbon/Unbabel-led) on EuroLLM, the from-scratch open European LLM covering all 24 EU official languages.
The second signature is Edoardo Ponti and Piotr Nawrot's inference-efficiency program, run jointly with NVIDIA: Dynamic Memory Compression (ICML 2024, up to ~4x LLM inference throughput via learned KV-cache compression), its DMS successor, and The Sparse Frontier, the largest empirical study of training-free sparse attention. Around it sit Mirella Lapata's long-context summarization line (MovieSum, discourse-driven planning) and Pasquale Minervini's group with MMLongBench (NeurIPS 2025 spotlight) and the SCOPE self-play open-endedness work.
Attribution discipline: the HPLT corpora are consortium artifacts and EuroLLM is Lisbon-led with Edinburgh as partner — Edinburgh is credited as a key contributor, not sole owner. The in-window picture is honest about gaps: multilingual data infrastructure, MT, and inference efficiency are strong; there is no in-window Edinburgh-owned scientific foundation model.
People
- Mirella Lapata Google Scholar — Professor, ILCC
- Edoardo M. Ponti Google Scholar — Assistant Professor, ILCC; Visiting Professor, NVIDIA
- Alexandra Birch Google Scholar — Associate Professor, Informatics (machine translation, multilingual LLMs, evaluation)
- Ivan Titov Google Scholar — Professor of NLP & ML, ILCC; dual appointment University of Amsterdam (ILLC)
- Amos Storkey Google Scholar — Professor of Machine Learning & AI; BayesWatch group