📣 Try out the NEW Gateway to Research and let us know what you think.

We're looking for users to test the new service during August and September and share their feedback. Express your interest by completing this short form.

AI HUB IN GENERATIVE MODELS

Lead Research Organisation: UNIVERSITY COLLEGE LONDON
Department Name: Computer Science

Abstract

Generative Models are AI models that can generate data. Recently researchers have shown that by training these models on large amounts of data (text data from the internet and images) these models learn to understand the regularities of our text and image world so well that they can generate responses to questions and create new images with surprising fidelity. This heralds a new era in which computers can assist humans to carry out tasks more efficiently than ever with significant opportunities for society, science and industry. However, these advances need significant research still -- how to make them train efficiently on different problems, how to understand their reliability and adherence to ethical norms.
 
Description The AI Hub in Generative Models has made major advances in its involvement with sovereign and trustworthy generative-AI capability. Two flagship programmes (UK-LLM and Nightingale AI) made substantial progress. UK-LLM (hub lead Prof. Pontus Stenetorp) expanded the UK's ability to build high-quality language-reasoning models for Welsh and other Celtic languages, addressing long-standing gaps in public-sector AI provision. Nightingale AI (hub lead Prof. Aldo Faisal) advanced the development of a multimodal health foundation model capable of integrating clinical notes, imaging, waveforms and population-scale data.

Methodological breakthroughs included DiffRatio, a new one-step diffusion-model training method that halves GPU memory requirements and enables ultra-fast image generation. Additional achievements included new evaluation frameworks for foundation models, advances in energy-based diffusion modelling, molecular-property prediction using diffusion models, and research into adversarial vulnerabilities in generative systems.
The Hub also strengthened the national research ecosystem through major events, workshops and a new Work-in-Progress initiative that launched 21 collaborative early-stage projects.
Exploitation Route The research outputs provide clear pathways for further development. Nightingale AI will be taken forward with clinical and health-data partners to support diagnostics, research and decision-support tools. UK-LLM will continue to expand through collaborations with universities in Wales, Scotland and Northern Ireland, with expected uptake across public-service applications.
DiffRatio and related methodological work are positioned for adoption by generative-AI companies and academic groups seeking more efficient training pipelines. Evaluation frameworks and robustness research will support regulators, industry and safety researchers. The Dataset Creation & Challenge programme and Work-in-Progress initiative will continue to generate new collaborations and translational opportunities.
International partnerships established through Data for Policy and the iDAT programme will support joint research, curriculum development and future workshops. Engagement with other UK AI hubs, ELLIS units and Prob_AI will continue to shape shared research priorities.
Sectors Other
URL https://www.genai.ac.uk/work-in-progress
 
Description The Hub's sovereign-AI work is generating clear societal and economic impact through two major programmes led by Prof P Stenetorp and Prof A Faisal, alongside wider ecosystem collaboration and international policy engagement. The Celtic-language LLM programme, led by Prof. P Stenetorp, is reducing long-standing digital inequities for Welsh, Scottish Gaelic and Irish speakers. By delivering a state-of-the-art Welsh model, creating new evaluation datasets and expanding partnerships across the UK, the work is enabling more inclusive access to public services, education and cultural resources. The Nightingale AI project, led by Prof. A Faisal, is transforming the UK's health-data ecosystem by building a sovereign, privacy-preserving multimodal foundation model capable of integrating clinical notes, imaging, ECGs, lab results and wearable data. This has the potential to improve diagnostics, accelerate biomedical research, support clinical decision-making and enable new digital-therapeutic tools, strengthening the UK's health-tech sector. These advances are reinforced by the Hub's Work-in-Progress initiative, which is catalysing early-stage collaborations across industry, academia and the public sector, and by international engagement through Data for Policy, which contribute to the global discussions on AI governance and responsible innovation. Through the hub's research outputs, methodological contributions such as DiffRatio, energy-based diffusion-model estimators and robust evaluation frameworks are helping to nucleate emerging areas in efficient generative modelling, trustworthy AI and multimodal health modelling. Workshops across biology, automation and creative AI have created new interdisciplinary communities, while the Work-in-Progress initiative has accelerated the formation of cross-institutional research teams. Public-sector bodies have engaged with Hub research through ongoing work on responsible-AI literacy frameworks (iDAT), which will support educators and policymakers. Private-sector organisations, including AstraZeneca, Nvidia, Bosch, Google Research and DeepMind, are collaborating on methodological advances, evaluation frameworks and domain-specific applications such as molecular property prediction. Through Prof. J Stilgoe's contributions, the hub contributed to a Privacy International report 'Nowhere to Hide' which has generated early societal impact by strengthening how the risks of AI-driven geolocation are understood and communicated across civil-society, policy and regulatory communities. The report shows that modern vision-language models can infer a person's location from a single image, creating new threats such as tracking, profiling, gender-based harms and commercial exploitation. Through its involvement, the Hub helped ensure these findings were framed not only as technical capabilities but as issues of public accountability, privacy and human rights. This has enabled Privacy International to present a clearer, more actionable analysis for policymakers, NGOs and regulators who rely on such evidence to shape debates on AI governance and surveillance. The report is now informing advocacy on safeguards for multimodal AI, raising public awareness of emerging privacy risks and supporting discussions on responsible deployment of generative and vision-language models, contributing to wider societal understanding and protection against new forms of digital harm.
First Year Of Impact 2026
Impact Types Societal

Policy & public services
 
Description DemystifAI Learning Community at UCL
Geographic Reach Local/Municipal/Regional 
Policy Influence Type Influenced training of practitioners or researchers
 
Description Privacy Risk and Policy Implications of AI Geolocation
Geographic Reach Multiple continents/international 
Policy Influence Type Participation in a guidance/advisory committee
URL https://www.privacyinternational.org/report/5736/nowhere-hide-privacy-risks-and-policy-implications-...
 
Description Automation Working Group workshop 
Form Of Engagement Activity Participation in an activity, workshop or similar
Part Of Official Scheme? No
Geographic Reach National
Primary Audience Other audiences
Results and Impact The Automation working group, led by Mike Wooldridge, convened a workshop 'Rethinking Multi Agent Systems'. With the emergence of large language models, interest in multi agent systems has surged, particularly in the potential for LLM powered agents capable of autonomous action and interaction. The workshop brought together researchers examining how LLMs can be used to realise the classic vision of multi agent systems and what established MAS principles can contribute in the LLM agent era. Discussions focused on structuring systems as collections of interacting agents, each playing a specific role, to support complex problem solving. There were 80 attendees from varying academic career levels as well as industry representation.
Year(s) Of Engagement Activity 2025
 
Description Creative AI & Multimodal Brainstorm - Gen AI 
Form Of Engagement Activity Participation in an activity, workshop or similar
Part Of Official Scheme? No
Geographic Reach National
Primary Audience Other audiences
Results and Impact The Creative AI and Multimodal Models working groups, led by Lourdes Agapito and Yukun Lai, hosted a workshop designed to foster collaboration across these communities. The event provided a dedicated space for members and affiliates to meet face to face, share emerging ideas, and develop potential research collaborations. Postdoctoral researchers were strongly encouraged to participate, reflecting the Hub's commitment to supporting early career researchers and strengthening interdisciplinary networks. 30 people attended from across the hubs network at various stages of their academic career.
Year(s) Of Engagement Activity 2025
 
Description Gen Biology Workshop 
Form Of Engagement Activity Participation in an activity, workshop or similar
Part Of Official Scheme? No
Geographic Reach National
Primary Audience Other audiences
Results and Impact Led by Prof. Magnus Rattray and Dr. Jason Hartford, this workshop brought together researchers working at the intersection of generative AI and biological discovery, creating a focused forum to explore how emerging models can support next-generation life-science research. Presentations covered key frontier areas including the use of generative models for genome design, the development of virtual cells and digital twins for personalised medicine, and large-scale efforts to map the human body at single-cell resolution to enable future chromosomal-level modelling. The event highlighted the rapidly expanding applications of generative AI in biology, from modelling cellular behaviour to designing biological system, and underscored the need for strong interdisciplinary collaboration between AI researchers, computational biologists and biomedical scientists. Discussions addressed both the opportunities and the technical and ethical challenges of applying generative models to complex biological questions. As a direct outcome, participants established monthly research meet-ups to maintain momentum, deepen collaboration and support the development of a growing UK community at the interface of AI and biological science.
Year(s) Of Engagement Activity 2025
 
Description Hub Launch 
Form Of Engagement Activity Participation in an activity, workshop or similar
Part Of Official Scheme? No
Geographic Reach National
Primary Audience Industry/Business
Results and Impact Hub Launch to the public. We had around 200 people from across industry and academia attend our vision for the Hub and how they can get involved.
Year(s) Of Engagement Activity 2025
URL https://www.genai.ac.uk/news-feed/gen-ai-hub-launches-at-the-crick