Fabio J. Fehr PhD

Dr. Fabio J Fehr is a Postdoctoral Researcher in machine learning at the University of Oxford, where he will join the Torr Vision Group. His research lies at the intersection of artificial intelligence, natural language processing, and agentic frameworks, with an emphasis on socially grounded applications of AI.

He completed a Master’s degree in Statistics at the University of Cape Town, focusing on biomedical modelling and deep learning, before pursuing a PhD in theoretical Machine Learning at EPFL and the Idiap Research Institute. His doctoral research examined the foundations of attention mechanisms and connected them to Bayesian nonparametric statistics in deep generative modelling. He also has industry experience at Amazon working on code generation, alongside entrepreneurial work in creative AI applications, including music generation. He is joining Oxford to work on AI for legal modelling, motivated by a belief that AI research carries a responsibility to address real societal challenges.

My research is grounded in statistical machine learning and deep learning, with a strong emphasis on tackling problems from first principles. I am drawn to problems where strong theoretical foundations matter, collaboration across disciplines is essential, and the resulting systems have the potential for meaningful societal impact.

Agentic and Retrieval Augmented Generation (RAG) models for legal reasoning: Legal reasoning is a text-heavy, complex, and high-stakes problem that remains largely unsolved. Legal NLP must contend with long documents, jurisdictional specificity, linguistic nuance, and the need for transparent and justified reasoning. This makes it an ideal domain for principled machine learning research that combines retrieval, grounded generation and multi-perspective reasoning. The field is new, inherently collaborative with clear relevance to both public institutions and industry.

Diffusion language models for long-form generation: Diffusion models offer a promising alternative to left-to-right generation by enabling global refinement and iterative improvement. This paradigm is particularly well suited to long-form and structured outputs such as legal documents, mathematical proofs, and code, where coherence, revision, and global consistency are more important than token-level prediction.

Modelling music and creativity with language models: This is a personal passion project. Music is a structured, sequential form of storytelling with rich temporal, harmonic, and timbral patterns. Modern language models are well positioned to capture this structure. By learning meaningful representations of musical elements such as rhythm, key, and timbre, we can build systems that both analyse and generate compelling music. Let’s jam on this!

• Agentic and Retrieval Augmented Generation (RAG) models for legal reasoning
• Diffusion language models for long-form generation
• Modelling music and creativity with language models