Home

Google Scholar • ArXiv • Twitter • Email

Dr. Kfir Dolev (Kuh-Fear)

AI Reasoning Post Doctoral Researcher

Tel-Aviv University

⚛️🎓💻ℹ️🔬🌌

AI researcher and theoretical physicist exploring information processing in physical systems.

Research Interests

Can we train general purpose LLMs mostly from synthetic data? What is the connection between the structure of this data and physical law?
What realistic approximations of the theoretically most powerful prediction agent (the Solomonoff inductor) exist? Does logical reasoning naturally emerge in good approximation schemes?

Academic History

I will soon begin a post doctorate working with Prof. Yoav Levine on fundamental models of AI reasoning.

In 2024-2025 I conducted mechanistic interpretability research with collaborators at SITP, in particular understanding how neural network superposition works, and how the networks are able to perform computation on superposed information.

In 2018-2024 I was a PhD student at the Stanford Institute for Theoretical Physics under the co-supervision of Ayfer Özgür and Patrick Hayden. My work was concerned with understanding how quantum information can move and be processed in spacetime, and searching for practical quantum information protocols inspired by the holographic principle.

In 2017-2018 I completed a masters program in theoretical physics at the Perimeter Institute. There I did research on entanglement in continuous variable quantum systems in the Barrio-RQI group, supervised by Eduardo Martin-Martinez.

In 2013-2017 I received a B.S. in physics, and a B.S. in computer science at the University of California, Santa Cruz. There I worked, under the supervision of Stefano Profumo on vacuum stability of Two-Higgs-Doublet models.

Selected Publications & Talks

A. Cowsik, K. Dolev, and A. Infanger, The Persian Rug: solving toy models of superposition using large-scale symmetries. arXiv preprint arXiv:2410.12101

We extract exactly the algorithm learned by the most well known model of neural network superposition, and distill it into a set of weights resembling a Persian rug, which matches the learned loss exactly.

Building a universal quantum computer in a simple QFT (QPV 2023 Talk)

I discuss how to build a universal quantum computer in a simple QFT through a series of embedded quasi-stable subspaces.

K. Dolev and S. Cree, Holography as a resource for non-local quantum computation. arXiv preprint arXiv:2210.13500

We show in holography that the boundary acts as a resource for non-local quantum computation, with the implemented computation describing the bulk dynamics.

K. Dolev, Constraining the doability of relativistic quantum tasks, preprint, arXiv:1909.05403

We show that a quantum channel acting collectively on systems located at any number of initial locations and distributing the result to any number of final locations can be accomplished without ever bringing the systems together. A recording of a talk is available here.

S. Cree, K. Dolev, V. Calvera, & D. J. Williamson, Gauging the bulk: generalized gauging maps and holographic codes, preprint, arXiv:2108.11402

We build a toy model of a holographic code with a boundary global symmetry dual to a bulk gauge symmetry.

Page updated

Google Sites

Report abuse