Dr Tianjin Huang

I am a Assistant Professor in Computer Science in the Department of Computer Science at the University of Exeter. I founded and currently lead the Reliable and Generalizable AI Laboratory (REGAL Lab). I am also a long-term Visiting Researcher at Eindhoven University of Technology (TU/e).

My research focuses on developing reliable and adaptive intelligence: AI systems that can learn robust representations, construct predictive models of the world, reason and act reliably, and adapt their computational effort to the complexity of each task. My work spans foundation models, world models, multimodal and agentic AI, and AI for science and biomedicine. A recurring objective is to make advanced AI systems more robust, generalisable, efficient and trustworthy, particularly when they operate with imperfect data, under uncertainty, or within limited computational budgets.

Through REGAL Lab, I lead research on the stable and resource-efficient training of large language models, reliable multimodal reasoning and decision-making, predictive world models for planning and action, and robust foundation models for scientific and medical applications. The laboratory brings together PhD students, master’s students and research interns working towards AI systems that can perform reliably in complex real-world environments.

Before joining Exeter, I was a University Researcher at Eindhoven University of Technology, where I worked on robust, frugal, resilient, secure and explainable machine learning. I was also a core researcher in the European Defence Fund–supported KOIOS project, investigating dependable machine-learning technologies for safety- and resource-critical settings. I continue to collaborate closely with researchers at TU/e through my visiting position.

My research has been published in leading artificial intelligence and machine-learning venues, including ICML, NeurIPS, ICLR, CVPR, MICCAI, Learning on Graphs, ECML PKDD, and Information Fusion. Recent contributions include SPAM, an optimiser for stable and memory-efficient large-language-model training, and GradientStabilizer, a lightweight approach to improving training stability across models and precision settings. I received the Best Paper Award at the inaugural Learning on Graphs Conference in 2022. My research has also received support from NVIDIA and UK national AI and high-performance-computing facilities.

REGAL Lab aims to develop AI systems that remain reliable in the imperfect real world, adapt computation to genuine task demands, and translate advances in learning, prediction and reasoning into scientific and societal benefit. My principal research areas include:

• Reliable and Efficient Foundation Models
• Robust and Generalisable Machine Learning
• Predictive World Models and Embodied Intelligence
• Multimodal Reasoning and Agentic AI
• AI for Science and Biomedicine

For further information about my research, publications and group, please visit my personal website.