I'm an MS student in Statistics at Stanford and a graduate researcher at MIT with Prof. Dimitris Bertsimas. Before this, I studied Computer Science and Economics at Bocconi and wrote my thesis at Harvard with Prof. Tomer Ullman. I also co-founded BAINSA, Italy's largest student-run AI research lab.

My research is on building intelligent systems that create new scientific knowledge, autonomously and safely. In particular:

• how to teach a language model "research taste" using signal from its own exploration;
• how to use test-time compute to explore the space of text-based solutions to open scientific problems, for example through evolutionary search or test-time learning;
• how to use these methods reliably on high-impact open problems in science and industrial R&D.

• Modeling the Object Representations Underlying Human Physical Reasoning. Andrey Gizdov*, Andrea Procopio*, Lorenzo Caputi*, Georgi Ivanov, Yichen Li, Daniel Harari, Tomer Ullman. Preprint.
• Do AI Indices Really Measure AI Activity? Riccardo Bertamini, Hugo Arsenio, Lorenzo Caputi, Andrea Procopio. Preprint.

* Equal contribution

• Probabilistic matrix factorization: seven ways to fill in a matrix with missing entries, compared on simulated data and on 100,000 movie ratings, with notes on the theory. Inspired by my research assistant work with Prof. Omiros Papaspiliopoulos at Bocconi.
• Rubik's cube solvers: five solvers written from scratch, from the beginner's method to the shortest possible solution and a network trained by reinforcement learning, compared on the same scrambled cubes.
• Telomere length and the limit of human lifespan: why the range of telomere lengths compatible with life narrows with age, and what that says about an upper limit to human lifespan. With Andrea Procopio and Lorenzo Battaglia.
• Simulated annealing on random 3-SAT: how close simulated annealing gets to the point where random 3-SAT formulas stop having solutions, and how that changes with formula size and annealing time.
• Football player valuation: predicting players' market values when a few stars dominate the error. With Andrea Procopio, Filippo Strub, Andrea Pettenon and Paride Lauretti.
• Benchmarking image upscalers: interpolation filters, open neural networks and online services, scored by their structural similarity to the original photo.
• Graph convolutional networks on citation networks: classifying papers by topic from their words and citations, reproducing Kipf and Welling on Cora, CiteSeer and PubMed.
• Also: strategies for Mastermind and hangman.

Website · Google Scholar · LinkedIn · Email