Live projects board
Hub researchers are collaborating on dozens of projects. Here are a selection, including some which need collaborators with specific expertise.
-
This project aims not only to learn representations, but also to develop generative models that use learning‑to‑structure to improve efficiency, generalisation, and the ability to express and estimate uncertainty.
-
Description:
Current Gen AI methods are behemoths, requiring internet-scale data, massive compute and energy resources to work. Their success derives strongly from the flexibility afforded by their size and the scale of data observed. But does this need to be the case?Recent work in representation learning shows judicious modelling constraints can be both effective and efficient by structuring observations to enhance reuse and compositionality---small models effective in low-resource settings.
This project aims to go beyond simply learning representations to develop generative variants that leverage learning-to-structure to be more efficient and effective, better amenable to generalisation, and the expression and estimation of uncertainties.
Who’s involved?
Sid Narayanaswamy, University of Edinburgh
-
This project is seeking collaborators
We are seeking collaborators who:Are interested in developing methodology and efficient modelling beyond LLMs and Diffusion Models.
Have interests aligned with human-like learning and cognitive science for features of representation learning and mechanistic interpretability.
Get in touch:
If you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
-
This project will develop models that can generate realistic cellular trajectories by training on single-cell time course data. We are also investigating the most suitable explainability frameworks for spatial omics datasets
-
Description:
One aspect of our project is the development of models that can generate realistic cellular trajectories by training on single-cell time course data. These models learn from snapshots, model time evolution, and capture both continuous variation and discrete variation into alternative cellular lineages. By capturing the dynamical and branching structure of the data, we hope to combine model interpretability with high generative accuracy. Another aspect of our project is to incorporate the spatial tissue context that is available from spatial omics assays, to better characterise diseased tissues. Here, one challenge is how best to deal effectively with multiple spatial modalities, while another challenge is to deal with large differences in spatial scales for cell and tissue interactions. We are also investigating the most suitable explainability frameworks for spatial omics datasets.Who’s involved?
Magnus Rattray, University of Manchester
David Barber, University College London
Yuhan Wang, University of Manchester
-
This project is seeking collaborators
We would be interested in collaborating with scientists and clinicians who are applying single-cell and spatial omics assays. We are also interested in collaborations on methodology development.Get in touch:
If you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
-
The project tackles implausible, misaligned outputs in multimodal generative models by analysing current model behaviour, developing cross‑modal knowledge‑aware architectures, and collaborating with industry to enable robust, practical applications.
-
Description:
Multimodal generative models enable content in diverse modalities to be generated, but also more specialised content for applications in e.g., animation, biology, healthcare, medical domains. While significant progress has been made, current generative models may still produce implausible results, results which are not expected or misaligned with the input modalities. High level reasoning or exploiting knowledge from different modalities remain challenging. By working together with hub members, industry and beyond, we aim to achieve the following: 1) Better understanding the behaviour of existing multimodal generative models, their capabilities and limitations, including more systematic performance evaluation. 2) Developing new multimodal generative models that effectively exploit the consistency and complementary prior knowledge in different modalities to improve capabilities of multimodal generative models. 3) Working with industries to develop techniques that lead to practical applications, in areas such as Creative AI.Who’s involved?
Yukun Lai, Cardiff University
Yuanbang Liang, Cardiff University
-
This project is seeking collaborators
Expertise in many domains (audio, NLP, knowledge representation) and real-world problems/insightsIf you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
-
This project develops a novel flexible memory system that enables AI assistants to maintain truly coherent, long-term conversations by preserving raw conversational data and supporting multi-resolution retrieval without information loss.
-
Description:
Current long-context LLMs suffer from context rot, where reasoning quality degrades as context windows expand. Existing memory-based solutions are inherently lossy, discarding portions of the original dialogue. To address these challenges, the system employs two foundational design principles. First, an indexed architecture ensures original conversation content remains accessible on demand rather than being compressed or abstracted away. Second, adaptive retrieval mechanisms enable access to information at varying levels of abstraction and granularity depending on task requirements. The integration of structured indexing with dynamic retrieval allows LLMs to perform reliable reasoning over extended conversational histories while maintaining complete access to the original interaction record, eliminating the information-fidelity trade offs inherent in current approaches.Who’s involved?
Mingtian Zhang, University College London
Pasquale Minervini, University of Edinburgh
-
This project is seeking collaborators
Collaborators are expected to contribute to:Conceptual development of the memory system design
Guidance on experimental design and evaluation methodology
Analysis of model behaviour under long-context settings Interpretation of results and co-authoring research outputs
Relevant expertise includes large language models, retrieval-augmented generation, reasoning evaluation, and experimental ML research.
Get in touch:
If you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
-
The research aims to build a data-driven recommendation system for CGHero, a marketplace of some 2,600 freelance CGI artists and processes, which can interpret and align multimodal creative briefs with artists’ styles and expertise.
-
Description:
CGHero operates a marketplace of some 2,600 freelance CGI artists and processes a large volume of client job requests that are inherently multimodal. Each request combines multiple forms of information. These diverse data sources capture how creative intent is communicated across text and visual media.The research aims to build a data-driven recommendation system that can interpret and align multimodal creative briefs with artists’ styles and expertise. Using generative and representation-learning models for creative media, the system will analyse text descriptions, visual references, and portfolio metadata to infer style, domain, and aesthetic compatibility, alongside practical constraints such as budget, availability, and client preferences. The system will generate a ranked shortlist of artists with interpretable explanations. The project will produce a publishable study and an implementable prototype for creative media matchmaking.
Who’s involved?
Mingfei Sun, University of Manchester
-
This project is seeking collaborators
A third party could contribute specialist expertise in multimodal machine learning, generative models for creative media, and human–AI interaction. Legal or ethics experts may be needed for data governance, privacy, and IP issues around creative content. Industry partners or design experts could support qualitative evaluation of artistic style matching and the interpretability of recommendations.If you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
-
This project proposes a unified research agenda that rethinks adversarial attacks in the age of generative models through a probabilistic and concept-level perspective.
-
Description:
We propose concept-based adversarial attacks, which generalize adversarial perturbations from single images to entire concepts represented as distributions. In addition, this project will revisit adversarial transferability from a probabilistic perspective. By analyzing transferability as an alignment problem between multiple victim distributions and a shared distance distribution, this work seeks to give a coherent theoretical account of existing methods and to guide the design of more principled, transferable attacks. Overall, this project argues that adversarial vulnerability in modern AI systems is fundamentally semantic and distributional, rather than purely geometric. By shifting adversarial attacks from images to concepts and from heuristics to probabilistic principles, this research will expose new failure modes of current models and inform the development of next-generation defenses for generative AI systems.Who’s involved?
Andi Zhang, University of Manchester
-
This project is seeking collaborators
The collaborator will provide expertise in generative modelling and large-scale experimental evaluation, contributing to model selection, benchmarking, and practical validation of the proposed methods.If you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
-
This project focuses on improving the efficiency and sustainability of multi‑modal generative AI systems, which currently rely on resource‑intensive latent diffusion models and encoder–decoder architectures.
-
Description:
Generative AI systems require substantial computational and energy resources, particularly in multi‑modal settings that involve multiple processing components. The research will investigate model compression, architectural refinement, and lightweight cross‑modal representations to reduce computational and memory demands while maintaining output quality. By reconsidering how components within the generative pipeline are designed and integrated, the project aims to lower energy consumption and environmental impact, supporting the development of scalable, high‑performance, and sustainable multi‑modal generative AI technologies.Who’s involved?
Arshdeep Singh, King’s College London
Mark D. Plumbley, King’s College London
Yukun Lai - Cardiff University
-
This project is seeking collaborators
On multi-modal generation, expertise is required on audio, visual and textual modality. On efficiency, expertise is required in model compression and optimization.If you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
-
This project will develop a responsible and critical AI literacy framework for educators and students to equip learners with the knowledge, skills and ethical awareness needed to navigate the opportunities and challenges presented by AI-driven technologies.
-
Description:
This project partners with the International Diagnostic and Admissions Test (iDAT) in Australia to develop a responsible and critical AI literacy framework for educators and students. Its aim is to equip learners with the knowledge, skills, and ethical awareness needed to navigate the opportunities and challenges presented by AI-driven technologies. By integrating technical understanding with sociotechnical perspectives, the framework will support more equitable, informed, and future-ready educational practices in the emerging age of AI.Who’s involved?
Canhui Lui, University College London
-
This project is seeking collaborators
We are seeking Expert Consultation on building up the feasible framework to assess the AI literacy.If you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
-
This two-stage training process, whilst pragmatic, raises some deep questions: why can't we train a one-step generator directly, without needing to first train a diffusion model?
-
Description:
Diffusion models have been hugely successful in generating high-quality data, e.g. images. The state-of-the-art pipeline is to first train a diffusion model, and afterwards distil this to a small number - possibly one - step generator. This two-stage training process, whilst pragmatic, raises some deep questions: why can't we train a one-step generator directly, without needing to first train a diffusion model? From the viewpoint of latent variable modelling, we propose to train a one-step generator from scratch, using a more sophisticated posterior variational approximation than a Gaussian, with the view that this may enable us to avoid the diffusion model completely.Who’s involved?
David Barber, University College London
Sid Narayanaswamy, University of Edinburgh
-
This project is seeking collaborators
The project requires experts in probabilistic modelling, with expertise in variational inference and approximate inference. The project also requires coding this (pytorch) to demonstrate results, and eventually to train on larger scale datasets.Industry is welcome to collaborate. We would value help with coding and running experiments; we would, of course, also welcome input into the methodology and potential application areas beyond image generation.
If you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
-
This project concentrates on the development of robust statistical approaches to evaluate the performance of foundation models, including LLMs and VLMs;
-
Description:
The project also concentrates on the development of approaches that are scalable, i.e. that do not rely on extensive annotated datasets.The project involves various elements: 1/ first, it involves the development of new statistical approaches that can leverage both small ground-truth human-annotated datasets coupled with large LLM-as-a-judge annotated ones; 2/ second, it involves benchmarking a number of foundation models using the proposed approaches; 3/ finally, it will also benchmark the proposed approaches against other baseline ones.
Who’s involved?
Miguel Rodrigues, University College London
-
This project is seeking collaborators
We would be primarily looking for collaborators using foundational models in various contexts. We are also open to collaborate with people with methodological backgrounds.Get in touch:
If you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
-
This project seeks to integrate a predictive model into Maia, Matillion’s AI-driven data engineering agent, to enhance the optimization of tool sequences.
-
Description:
This model will predict potential failures in tool sequences and recommend configurations aimed at preventing inefficiencies and errors. The University of Manchester, supported by Matillion funding, will develop this model leveraging their expertise in predictive analytics.Who’s involved?
Mingfei Sun, University of Manchester
-
This project is seeking collaborators
A third party could contribute expertise in predictive analytics, agentic AI systems, and large-scale system monitoring. Additional support from experts in software reliability, MLOps, and data governance would strengthen deployment, evaluation, and compliance.If you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
-
This project focuses on understanding and evaluating robustness in vision language models (VLMs) across a range of realistic visual perturbations and task settings.
-
Description:
VLMs are increasingly applied in real-world settings where visual inputs may be noisy, distorted, or incomplete. This project studies the effects of natural and digital image transformations, including noise, blur, and weather effects, applied at controlled severity levels across established multimodal benchmarks. The project will produce an open and extensible robustness evaluation and analysis toolkit that collaborators can apply to their own models and datasets. The resulting insights will support model development, fine-tuning, and deployment decisions, and contribute to a clearer understanding of robustness properties in modern vision–language systems. Overall, the work aims to provide practical evaluation methods and empirical evidence to support more reliable and trustworthy multimodal AI.Who’s involved?
Rohit Saxena, University of Edinburgh
Pasquale Minervini, University of Edinburgh
-
This project is seeking collaborators
Experience in Vision Language Models training and evaluation; robustness and adversarial examples in vision and text; engineering and research inputs.If you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
-
Building on a successful UK-LLM, Bangor University and Nvidia collaboration producing state-of-the-art Welsh language models, this project aims to expand LLM capability across all Celtic languages.
-
Description:
Celtic language users face significant disadvantages as established LLM technology performs substantially worse for their languages compared to English. The 2025 Welsh collaboration produced a new state-of-the-art model alongside three cutting-edge evaluation datasets. Welsh, as the most well-resourced Celtic language, served as a reasonable first step, but expansion to Scottish Gaelic and Irish is planned through extended partnerships with the University of Edinburgh and Queen's University Belfast. A research associate fluent in at least one Celtic language will be recruited to serve as the interface between academic institutions and help develop next-generation Celtic LLMs. Additionally, a large cohort of annotators will be hired to create full translations of established benchmarks such as Global MMLU, ensuring robust evaluation coverage and presence in international benchmarks used by LLM developers globally.Who’s involved?
Pontus Lars Erik Saito Stenetorp, University College London
-
This project is seeking collaborators
Collaborators are expected to have knowledge of Celtic languages, Celtic language technology, public and private deployment of Celtic language technology, and/or cutting-edge development of LLMs.If you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
-
This project builds an agentic pipeline for automated mechanistic interpretability, converting model behaviour analysis into verified, targeted interventions — with a primary application in LLM safety monitoring.
-
Description:
Current mechanistic interpretability research is largely manual and explanatory, rarely producing actionable outcomes. This project operationalises the field as an end-to-end agentic pipeline combining an investigator module — responsible for activation-to-text decoding — with tool-calling agents that localise features and circuits, generate discriminating counterfactual inputs, and test competing mechanistic hypotheses systematically. The multi-turn, tool-mediated architecture avoids reliance on a model's own natural-language rationalisations, enabling more rigorous causal attribution. Outputs are domain-grounded, actionable reports paired with minimal-change interventions verified against fix-rate, side-effect, and jailbreak-robustness metrics. The flagship application is safety: a guardrail model with privileged activation access can detect dangerous intent or knowledge composition in a target LLM prior to verbalisation, triggering graded control responses — addressing a critical gap between interpretability research and deployable alignment tooling.Who’s involved?
Pasquale Minervini, University of Edinburgh
Fazl Barez,
Yihong Chen
-
This project is seeking collaborators
We are seeking collaborators with (a) access to frontier or high‑impact deployed models and safety evaluation pipelines, and/or (b) expertise in mechanistic interpretability tooling (SAEs, causal tracing) to co-develop and stress-test the system on real-world failure modes.NLP, AI Safety, and Mechanistic Interpretability are all fundamental skills for this project.
If you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
-
This project will develop a synthetic benchmark to systematically analyse the capabilities of models along different axes of difficulty and used to study the limitations of existing model architectures.
-
Description:
LLM-based reasoning models are difficult to apply when inference rules and statistical patterns need to be learned from data that is outside the training corpus of the LLM (e.g. proprietary data) or that is only present in small amounts. We therefore need models that can learn to reason in a robust and systematic way from limited amounts of data. This project will first develop a synthetic benchmark to enable systematic analysis of the capabilities of models along different axes of difficulty. Using this benchmark, we will study the limitations of existing model architectures and develop novel approaches.Who’s involved?
Steven Schockaert, Cardiff University
-
This project is seeking collaborators
I am looking for university partners with expertise/interests in neurosymbolic AI, multimodal reasoning, and/or representation learning. I am looking for industry partners who can provide use cases and/or relevant datasets.If you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
-
This project proposes a generative framework that synthesises realistic data while targeting high‑uncertainty regions critical for early‑warning performance in disease detection.
-
Description:
This project addresses the challenge of early disease detection from sequential signals, where progress is limited by scarce data, restricted population diversity, and strict privacy constraints. We propose a causally informed, physiologically grounded generative framework that synthesises realistic data while targeting high‑uncertainty regions critical for early‑warning performance. Evaluation on public ICU physiological time‑series and respiratory‑sound datasets will assess gains in accuracy, robustness, and physiological plausibility. The work aims to establish a trustworthy generative foundation that strengthens remote‑monitoring capability and supports UKRI priorities in health resilience, data‑driven innovation, and safe, reliable AI.Who’s involved?
Yuhua Li, Cardiff University
-
This project is seeking collaborators
Healthcare providers: contributing data and clinical insightIf you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
-
Self-play has achieved remarkable success in two-player games, but these approaches remain restricted to domains with well-defined answers. This project aims to extend self-play to open-ended domains, enabling models to explore and self-improve across diverse areas.
-
Description:
Reinforcement learning with verifiable rewards (RLVR) has driven recent advances in LLM reasoning. However, it requires extensive human-annotated data for reward signals, which bottlenecks scalability and potentially limits performance to human-level. Self-play has achieved remarkable success in two-player games such as Go and chess. More recently, researchers have adapted self-play techniques to improve LLM reasoning without external data, showing promising results. However, these approaches remain restricted to domains with well-defined answers, where self-consistency can estimate question legitimacy. Extending self-play to domains with open-ended or long-form answers remains an open challenge. This project aims to extend self-play to open-ended domains, enabling models to explore and self-improve across diverse areas. We plan to address two core challenges: (1) generating high-quality open-ended tasks during self-play to guide learning, and (2) distinguishing and rewarding high-quality open-ended responses.Who’s involved?
Pasquale Minervini, University of Edinburgh
Cyrus Kwan
-
This project is seeking collaborators
Experience with RL (especially via self-play), NLP, Open-Ended Learning, and Continual Learning would be immensely useful for the project.If you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
-
Nightingale AI is building a sovereign, open, multimodal generative health foundation model trained on real biomedical, clinical and population data to accelerate research, improve clinical decision support, and power diagnostics.
-
Description:
The architecture employs a Perceiver-style design enabling asynchronous multimodal integration across heterogeneous health data—clinical notes, labs, vitals, ECG waveforms, imaging and wearables—arriving at different time scales. Rather than forcing inputs into a synchronized super-tensor, the model uses a shared latent space that selectively attends to whatever data is available at inference time. This modular approach allows new data streams or updated modality-specific encoders to be incorporated without retraining the entire foundation model. Currently functioning modalities include tabular Electronic Healthcare Records, ECG traces, clinical notes and X-ray imaging. The roadmap targets additional imaging modalities, genomics and wearable data, leveraging the UK Biobank. Nightingale AI is not language-based; it is a privacy-preserving engine designed to transform fragmented health data into actionable insights for discovery, care, system operations and digital therapeutics.Who’s involved?
Aldo Faisal, Imperial College London
Marek Rei, Imperial College London
-
This project is seeking collaborators
For academic and scientific collaborators, we’re seeking expertise in generative AI methods, biomedical image understanding, and multimodal omics. For clinical partners, we’re looking to co-develop and evaluate real-world use cases powered by Nightingale AI, grounded in the underpinning datasets that can be analysed within secure data environments. For industry partners, we’re keen to work with their proprietary data and to establish jointly sponsored “driver projects” that accelerate the development and deployment of GenAI applications in their sector.If you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
-
We are developing novel methods for inverse problems based on diffusion and flow models.
-
Description:
Inverse problems underpin a range of imaging and sensing applications, including image super resolution and denoising, weather data assimilation and reconstructing astronomical & medical images from limited measurements. This problem of inferring structure from limited observations is typically framed as Bayesian posterior computation, combining a prior (representing structural knowledge of desired outputs) with a likelihood (representing observations).There has been progress using GenAI, like diffusion models & flow maps, for inverse problems. The GenAI models act as data-derived priors, while observational information is included via guidance towards samples consistent with it. However existing works have been based on severe approximations leading to significant fine-tuning and ad-hoc improvements to work well. We are investigating better approaches that bypass these approximations, leading to improved fidelity and computational efficiency.
Who’s involved?
Yee Whye Teh
-
This project is seeking collaborators
I would be happy to collaborate with partners through the GenAI hub, particularly partners with expertise in the relevant scientific fields where inverse problems are found (e.g. astronomy, medical imaging, remote sensing, seismology).If you can offer the expertise needed for this project and would like to collaborate, please email us
Contact us
Collaborators sought
The following projects would welcome collaborators with the specific expertise described in the summaries below. If you have this expertise, please get in touch using the links provided and we’ll review your expression of interest for suitability.
We will only respond to expressions of interest that directly align with the project requirements. If you would like to get involved in hub activities but don’t have the expertise required by these projects, please see our FAQs page for more details.
-
This project develops efficient, reliable training methods for one-step generative models, enabling higher-quality image generation at reduced hardware costs through simplified frameworks that provide more stable, less biased learning signals.
-
Description:
This project aims to develop more efficient and reliable training methods for fast generative models, enabling higher-quality image generation with one-step models at reduced hardware cost. Existing approaches often rely on indirect or approximate training signals, which can introduce systematic bias, increase GPU memory requirements, and limit the quality of one-step generated outputs.The project will propose a simplified training framework that provides a more stable and less biased learning signal. By improving training efficiency and reliability, the proposed approach seeks to lower GPU memory usage while consistently producing higher-quality images.
Who’s involved?
Mingtian Zhang, University College London
Jose Miguel Hernandez Lobato, University of Cambridge
David Barber, University College London
-
This project is not currently seeking collaborators
-
This project aims to uncover the underlying mechanisms of the phenomena of diverse forms of forgetting
-
Description:
This study would include catastrophic forgetting, intentional unlearning, and time-dependent forgetting, across different learning algorithms by developing a unified analytical framework, ultimately enabling safer, more efficient, and more capable AI systems.
The researchers aim to uncover the underlying mechanisms of continual learning-related phenomena across different learning algorithms by developing a unified analytical framework, ultimately enabling safer, more efficient, and more capable AI systems.
Who’s involved?
Yarin Gal, University of Oxford
-
This project is not currently seeking collaborators
Other active projects
The researchers leading these projects are not currently looking for collaborators.