Alessandro Favero

I’m an inaugural Physics-AI Fellow at the University of Cambridge and the Sansom Research Associate at Emmanuel College. I’m also part of the interpretability team at Polymathic AI.

I work on the physics of learning and intelligence. My research combines theory and empirics to advance our scientific understanding of AI, with a focus on the representations deep neural networks acquire – of language, images, scientific data, and tasks. Recent topics include:

• memorization, generalization, and compositionality in generative and world models
• weight-space geometry, editing, and interpretability of foundation models
• formal grammars, language structure and learnability

Before Cambridge I did my PhD at EPFL with Matthieu Wyart and Pascal Frossard, with a summer at Amazon’s AI Labs in Stefano Soatto’s group working on multimodal LLMs. Earlier, I studied physics in Turin, Trieste, and Paris.

I’m always happy to hear from prospective students interested in the science of deep learning. Email volume is high, so a short note explaining your background and what specifically draws you to my work helps me get back faster.