Representational AI Lab

Learning AI systems that adapt, align, and evolve reliably.

RepsAI develops machine learning methods for building adaptive, aligned, and trustworthy AI systems through representation learning, generative modeling, optimization, continual learning, and unlearning.

Publications

About the Lab

The Representational AI Lab (RepsAI) focuses on machine learning methods for building AI systems that can adapt to new data and tasks, align with desired representations and objectives, and update their knowledge in a controlled and trustworthy way.

Our work sits at the intersection of representation learning, generative modeling, optimization, continual learning, and agentic AI. We are particularly interested in how representations can make modern AI systems more robust, controllable, efficient, and trustworthy.

Applications of our work include inverse problems, scientific AI, computational pathology, foundation model adaptation, diffusion and flow-based generative models, and LLM agents.

Research Themes

🧠

Representation Learning

Learning structured, robust, and transferable representations that support adaptation, alignment, and reliable generalization.

✨

Generative Models

Diffusion, flow-based, and latent-variable models for inverse problems, scientific AI, and controllable generation.

⚙️

Optimization & Inference

Optimization, probabilistic inference, and theory for scalable, efficient, and reliable learning systems.

🤖

Agentic & Continual AI

Adaptation, alignment, post-training, unlearning, and continual-learning methods for foundation models and LLM agents.

Current Members

PhD Students

PhD Student

Loukas Sfountouris

2024 – present

Deep Generative Models

Main supervisor: Paris Giampouras · Co-supervisor: Prof. Theo Damoulas

Website Email

PhD Student

Xietao Wang Li

2024 – present

AI/ML for Science

Co-supervisor: Paris Giampouras · Main supervisor: Prof. Tom Montenegro-Johnson

Website Email