RadEdit is a latent diffusion model trained to generate and edit chest X-rays from medical reports. It is described in detail in RadEdit: Stress-Testing Biomedical Vision Models via Diffusion Image Editing (F. Pérez-García, S. Bond-Taylor, et al., 2024).
“Agentic AI Research & Innovation enables bold, ambitious ideas that will advance our understanding of the foundations of agentic AI, foster real-world innovation, and embed safety and responsibility at the heart of AI Agent development." AI continues to evolve at an unprecedented pace, unlocking transformative…
This project aims to push the boundaries of agentic AI by addressing three critical challenges: long-horizon memory, safe and aligned tool usage, and adaptive reasoning. Current language models excel at text generation but lack agency—the ability to plan, act, and adapt in dynamic environments. The…
This project explores how Generative AI can transform supply chain management from rule-based automation to true autonomy. Building on the MIT autonomous supply chain testbed, it integrates multiple AI agents that learn, adapt, and coordinate decisions across forecasting, inventory, and replenishment—cutting costs by up to…
This project develops a unified framework for physically grounded world modelling that combines video-based temporal prediction with Gaussian Splatting for photorealistic 3D representation. A Physics Vision-Language Model translates natural-language instructions into transformations that respect physical constraints, enabling interpretable and goal-directed control in dynamic scenes. By…
This project introduces AgentGuard, a monitoring and routing system designed to improve reliability and cost-efficiency in Azure-based agent workflows. By analysing early trajectory signals—such as reasoning patterns and tool usage—within the first 10–30% of an agent’s execution, AgentGuard predicts task outcomes and dynamically halts or…
This project addresses the challenge of enabling AI agents to operate effectively in complex, realistic environments such as web navigation, computer use, and mobile interfaces. While current models excel in structured domains like mathematics and coding, they struggle with tasks requiring long-horizon planning, exploration, and…
Human intelligence is defined by the interplay of semantic and episodic memory. AI research almost exclusively develops semantic memory systems, but episodic memory has many open challenges, especially for vision. Episodic memory is the ability to recall specific past experiences and use the information together…
This project tackles the challenge of cultural misalignment in AI agents, which often leads to unsafe or inappropriate behaviour in diverse, real-world interactions. CuRA introduces a dynamic routing framework that leverages specialised cultural adapters to generate context-aware, low-risk responses across multiple turns and modalities. Paired…
This project focuses on creating robotic systems that can adapt and improve during deployment in dynamic, unstructured environments such as warehouses and industrial sites. It combines the reliability of classical robotics with the flexibility of learned policies, enabling robots to reason, plan, and recover from…
This project introduces a novel architecture for agentic AI systems that ensures accuracy, efficiency, and safety during reasoning. It addresses two key challenges—lack of steerability and absence of verifiable guarantees—by developing verifiers that can interject at any point in a model’s generation process. An auxiliary…
Towards The Psychological Security of Agentic AI This project addresses the critical challenge of auditing psychological safety in agentic AI systems designed to deliver evidence-based interventions for mental and cognitive well-being. While agentic AI chatbots can emulate psychotherapeutic interventions and enable personalized support, they also…
