Jonathan Richard Schwarz

ICL Visiting Professor · Imperial College London

TR Head of AI Research · Thomson Reuters

OII Research Associate · Oxford Internet Institute

AISI Expert Advisor · UK AI Security Institute

Research focus

My research targets efficient, general, and robust machine learning — particularly algorithms that transfer knowledge across tasks, letting agents build skill repertoires that accelerate future learning. A central goal is designing systems that abstract knowledge from related problems and reuse it effectively, enabling faster and more flexible learning on new tasks. Beyond foundational machine learning, I have a keen interest in the societal impacts of AI, including AI Safety, Economics, and Political Science. A particular focus is effective pluralism in the values and governance of Foundation Models, and ensuring shared stake in their societal consequences.

Previously, I was a Research Fellow at Harvard University and Co-Founder and Chief Scientific Officer at Safe Sign Technologies, a start-up focused on developing Foundation Models for professional work (acquired by Thomson Reuters in 2024). Before Entrepreneurship, I spent seven years at Google DeepMind as a Senior Research Scientist. During my time at DeepMind, I completed my PhD through the DeepMind–UCL programme, advised by Yee Whye Teh and Peter Latham. My PhD focused on sparse parameterisations and knowledge transfer for efficient ML, researching many ideas (Continual & Meta-Learning, Sparsity, Compression) that became pivotal in the development of large-scale Foundation Models. Before post-graduate work, I graduated top of class from the University of Edinburgh with a specialism in Machine Learning.

Research Interests

Large Language Models Societal Impacts AI Safety Data-centric Machine Learning Reasoning & Agents Sparsity Continual Learning Neural Data Compression Meta-Learning

News

June 2026 Excited to join the Oxford Internet Institute as a Research Associate.
May 2026 Pleased to join an inspiration exchange on Causal Reasoning & LLMs at The University of Cambridge.
May 2026 Very proud to officially launch Imperial's new Frontier AI Research Lab.
May 2026 Serving as Area Chair for the NeurIPS 2026 Evaluations and Datasets Track.
April 2026 New paper on measuring principal hierarchies under high-stakes competing demands.
February 2026 New paper on measuring Forgetting in LLM Post-Training.
February 2026 New paper on using aligning Language Model benchmarks with pairwise preferences.
January 2026 Serving as Area Chair for ICLR 2026.
January 2026 New paper on using Knowledge Graphs to construct LLM training data.
Nov 2025 Invited talks at Allen Institute for AI and University of Oxford.
Nov 2025 ADMIRE-BayesOpt accepted to TMLR.
Oct 2025 Scales++ released — efficient evals through data pruning.
Sep 2025 Invited panel discussion at The Royal Society on AI and the Law.
May 2025 Sparse Interpolated MoE accepted to ACL 2025 (Oral).
Jan 2025 Composable Interventions for LLMs accepted to ICLR 2025.
Dec 2024 Organised the NeurIPS 2024 workshop on Compositional Learning.
Sep 2024 CoLoR-Filter and MAC both accepted to NeurIPS 2024.
Aug 2024 Safe Sign Technologies acquired by Thomson Reuters.
Oct 2024 Paper on AI agents for scientific discovery accepted at Cell.
Apr 2024 Announcing the Harvard Efficient ML Seminar Series.
Jan 2024 Invited talk at the Stanford Information Theory Forum on neural data compression.