Research interests
What if ideas got taken seriously enough to see them through?
by day
Generative AI,
robotics, and
machine learning. The work is
architectural: how such systems are designed, shipped, and
kept running at scale to reach users. In production,
resource trade-offs like correctness versus latency
dominate.
by night
The information-theoretic foundations of
quantum mechanics. Existence is
distinguishability. A physical state is its pattern of
distinctions against all other states. From finite
precision, finite-dimensional quantum mechanics follows. The
core is formalized in Lean 4.
Guiding principle: Theory → Production.
Currently: submitting Existence as Distinguishability
to Physical Review X; follow-up work on its physical consequences
and on generalizing the kernel framework to pure mathematics; and a
methodology paper in which a Lean formalization, a semantic graph,
and a LaTeX manuscript jointly form the primary research artifact.
The published article becomes a rendering of the work, not its source.
Bio
I am a researcher and engineer based in Zurich. In May 2026 I am
joining Flexion Robotics as an
Applied Scientist. From 2023 to March 2026 I was an Applied
Scientist II at AWS in New York, working on generative AI:
hallucination detection for Q Business, agentic
natural-language data analysis for Quick Suite,
and enterprise retrieval systems.
I earned my PhD in 2022 from ETH Zurich, advised by
Prof. Emilio Frazzoli
and Prof. Stefano Soatto,
on plasticity, invariance, and priors in deep neural networks.
Before the PhD I spent six months as a research scientist at IDSIA under
Prof. Jürgen Schmidhuber
(co-inventor of LSTM). MSc Robotics, ETH 2016, Excellence Scholarship;
BSc Mechanical Engineering, ETH 2013, Outstanding Bachelor Award.
Beyond day-to-day engineering I maintain an active research line on the
information-theoretic foundations of quantum mechanics, deriving the
Schrödinger equation and the Born rule from a single ontological
principle, and on the corresponding Lean 4 formalization. My
published work has accumulated over
1000 citations across ICML, NeurIPS, and ICRA venues.
Selected work
Existence as Distinguishability: Quantum Mechanics from Finite Graded Equality
Julian G. Zilly ·
arXiv:2603.11900 ·
Lean 4 formalization ·
2026
A derivation of finite-dimensional quantum mechanics from a single
ontological principle: existence is constituted by distinguishability.
Comparison has finite resolution, distinctions are graded by a kernel
$K(x,y) \in [0,1]$ replacing binary equality, and two axioms
(finite capacity with saturation; universal relationality) determine
the state space $\mathbb{C}P^{N-1}$, complex coefficients, the
Born rule, unitary dynamics, and tensor products uniquely.
Standard QM is the $N\!\to\!\infty$ limit; finite $N$ is a natural
UV cutoff. The core mathematical derivation is machine-checked in
Lean 4 with zero sorry steps.
On Plasticity, Invariance, and Mutually Frozen Weights in Sequential Task Learning
Julian G. Zilly, Alessandro Achille, Andrea Censi, Emilio Frazzoli ·
NeurIPS 2021 ·
Scholar
Pre-training does not always help downstream tasks. We introduce the
notion of mutually frozen weights: weight configurations
where gradients on one task are annihilated by the geometry imposed
by another, producing a hidden form of plasticity loss. The work
characterizes when and why sequential task structure induces
representation rigidity, giving a concrete answer to the plasticity
question that underlies continual and transfer learning.
Recurrent Highway Networks
Julian G. Zilly, Rupesh Kumar Srivastava, Jan Koutník, Jürgen Schmidhuber ·
arXiv:1607.03474 ·
ICML 2017 ·
NVIDIA V100 award (Srivastava) ·
590+ citations
A recurrent architecture that extends LSTM with deep
transition functions between successive timesteps, allowing
step-wise computational depth without step-count-dependent vanishing
gradients. Establishes new state-of-the-art language-modeling
results on Penn Treebank and Hutter Prize Wikipedia, and introduces a
GCU-style analysis of gradient flow across compositional transitions.
The AI Driving Olympics at NeurIPS 2018
Julian G. Zilly, Jacopo Tani, Breandan Considine, Bhairav Mehta,
Andrea F. Daniele, Manfred Diaz, Gianmarco Bernasconi, Claudio Ruch, Jan Hakenberg,
Florian Golemo, A. Kirsten Bowser, Matthew R. Walter, Ruslan Hristov, Sunil Mallya,
Emilio Frazzoli, Andrea Censi, Liam Paull ·
arXiv:1903.02503 ·
NeurIPS'18 Competition Track, 2019
The first AI Driving Olympics (AIDO), a reproducible autonomous-driving
competition built on the Duckietown
platform. Tasks of increasing complexity, from lane following through
control with multiple vehicles and navigation with fleet management, tie together a
perception-planning-control pipeline in a benchmarkable form.
100+ international teams participated across the NeurIPS and ICRA editions.
Earlier work on glaucoma detection via entropy sampling
(best paper award, Computerized Medical Imaging and Graphics, 2017),
and the PhD dissertation Plasticity, Invariance, and Priors in Deep Neural Networks
(ETH Zurich, 2022). Full list on
Google Scholar.
