Mabrok Lab

Each recent paper in a few paragraphs: what it actually does, what it found, and where it fell short. No background assumed, and no numbers softened.

Research explained

A paper is written for the few hundred people who already work on the problem. That is what it is for, and it is why it reads like a wall to everybody else, including many of the people the work is meant to reach: clinicians, engineers, students, and the taxpayer who paid for it.

So here is the other version, grouped by what the work is about. Every number in it is a number from the paper. Where something did not work, it says so.

AI for healthcare

Borrowing the mathematics of a swing

HamVision: Hamiltonian dynamics as inductive bias for medical image analysisMedical Image Analysis, 2026

one scan a damped oscillator position · momentum · energy where the effort went no outline was ever supplied

Start here

Push a child on a swing and you are using one of the oldest equations in physics. The swing has a rhythm of its own, it stores the energy you give it and hands it back, and friction slowly settles it. Three hundred years of mathematics describe that behaviour exactly.

What we did

We built that swing into a neural network. Every image the model looks at has to pass through a narrow point in the middle of the network, and at that point we put a damped oscillator: a small piece of physics with position, momentum and energy. Instead of the model having to learn from nothing how information ought to move, it starts with a physical habit already in it.

What happened

It needs four times fewer parameters and seven times fewer arithmetic operations than the best previous model, and still leads or matches it on eight of nine medical imaging benchmarks, across nine different kinds of image. On skin lesions the margin is nearly 18 percentage points. On the smallest benchmark, 546 training images, it loses, which is what you would expect of a model that has structure to fill in.

Why it matters

A model this small can run on the hardware a hospital already owns, which is the difference between a result and a tool.

The part worth pausing on

The oscillator's energy map, which is just a picture of where the model is spending effort, settles on the part of the image that decides the diagnosis. Nobody told it where to look. It was never given a single outline to copy.

Asking a neural network the question you would ask a spacecraft

Controllability Analysis for Vision State Space Models: A Structural Interpretability FrameworkIEEE Transactions on Neural Networks and Learning Systems, 2026

thrusters can I steer it? ≈ early layer middle deep layer depth which patches can steer the state?

Start here

Before anybody flies a spacecraft, an engineer asks a blunt question: with the thrusters I have, can I actually steer this thing wherever I need it to go? The mathematics that answers it is called controllability, and it has been the first question in control engineering since the 1960s.

What we did

We asked a neural network the same question. Which parts of the picture can actually steer what is going on inside the model? Whatever can steer it is, by definition, what the model is paying attention to.

What happened

It beats Grad-CAM, the tool everyone currently uses to see what a vision model is looking at, on six of seven datasets. It needs one pass through the network, no retraining, and no changes to the model. Watching it layer by layer, the attention starts as a vague wash over the whole image and sharpens with depth onto the structures a radiologist would name.

Why it matters

"Show me what the model looked at" is the first thing a clinician asks and the hardest thing to answer honestly. Most answers are plausible pictures. This one is a measurement, with a definition behind it that predates the model by sixty years.

The longer version, with the method set out in full, is X-VMamba.

The reason medical AI fails in hospitals is usually not the model

MammoClean: Toward Reproducible and Bias-Aware AI in Mammography Through Dataset HarmonizationIEEE Access, 2026

one machine another a third its own words its own brightness its own population one shape, one vocabulary breast density, two datasets how far apart they really are

Start here

Two hospitals photograph the same breast and get different pictures. Different machine, different settings, different brightness, different patients, different words in the file. Train a model on one hospital's pictures and it can quietly learn the hospital rather than the cancer.

What we did

Rather than another model, a tool. MammoClean takes the large public mammography datasets and puts them into one honest shape: the same case selection, the same handling of left and right, the same intensity correction, the same words for the same things. Then it measures what was different in the first place.

What happened

Across three of the big public datasets, the differences are not small: breast density and how often abnormalities appear both shift substantially from one to the next. Models trained on the corrupted versions measurably lose accuracy against the same models trained on the curated ones.

Why it matters

This is the unglamorous half of the field, and it is where most of the failures actually live. A result that only holds on the dataset it was trained on is not a result a hospital can use.

The code is on GitHub, so anyone can check the claim or use it.

Agentic AI

What a language model means, and what it is able to say

Latent Semantic Manifolds in Large Language ModelsPreprint, 2026

one territory, one word on the boundary between two words every shade of meaning, continuously

Start here

A language model thinks in a space with no edges: every shade of meaning is available to it, continuously. Then it has to speak, and speaking means choosing from a fixed list of a few tens of thousands of tokens. It is a painter with every colour in existence, handed a box of crayons.

What we did

We measured the gap. Treat the model's internal states as points on a curved surface, and the vocabulary carves that surface into territories, one per token. Every meaning inside a territory comes out as the same word. How much is lost doing that is a geometry question, and we proved two theorems about it.

What happened

The predictions hold across six models from 124 million to 1.5 billion parameters. The gap grows in a straight line with size, closely enough to be called a law. And there is a stubborn core of meanings that sit right on the boundary between two words, which does not shrink as the models grow.

Why it matters

Some of what a language model gets wrong is not a bug in the training and will not be fixed by a bigger model. It is the cost of having to speak in words at all. That is a different kind of limit, and it is worth knowing which kind you are up against.

Robust and optimal control

Why a tall crane does not shake itself to pieces

Negative Imaginary Systems Theory for Nonlinear Systems: A Dissipativity ApproachIEEE Transactions on Automatic Control, 2025

+ − P(s) Q(s) u y NEGATIVE IMAGINARY STRICTLY NEGATIVE IMAGINARY λmax P(0)Q(0) < 1 is necessary and sufficient for internal stability

Start here

Anything long and flexible rings like a bell: a crane, a robot arm, a satellite's solar panel, a tall building. Push it at the wrong rhythm and the wobble grows instead of dying away. The engineer's job is to guarantee, before it is built, that it will settle.

What we did

There is a piece of control theory called negative imaginary systems theory that gives exactly that guarantee, and gives it without needing to know the structure's numbers precisely, which is what makes it usable on a real machine. Until now it only worked for linear systems: the tidy, idealised version. We extended it to nonlinear systems, which is what everything actually is.

What happened

The guarantee now covers two flexible systems connected to each other, which is where the surprises usually come from, and it comes with a worked control design for a nonlinear mass and spring.

Why it matters

A guarantee that only holds for the idealised version is a guarantee you cannot lean on when the thing is heavy and moving fast.

Dynamics and game theory

A crowd changing its mind obeys the same law as a wobbling bridge

Passivity Analysis of Replicator Dynamics and Its VariationsIEEE Transactions on Automatic Control, 2021

strategy A B C it settles = a population a flexible structure negative imaginary the same property, both times

Start here

People copy what seems to be working: a route to work, a way of trading, a habit. Strategies spread through a population and compete. Sometimes the population settles on one. Sometimes it circles for ever.

What we did

We treated the population as a machine with an input and an output and asked a control engineer's question about it: is it passive? A passive system can absorb what is pushed into it but cannot manufacture energy of its own, and passive things wired together are stable. That makes it a way of proving a population will settle, rather than simulating it and hoping.

What happened

The standard model of strategies spreading turns out to be passive in the strictest sense, so the stability follows. A richer version, where the population has momentum and can overshoot, is not passive. It has a different property instead, and that property is negative imaginary: the same one that keeps the crane above from shaking itself apart.

Why it matters

It is the same mathematics. A vibrating structure and a population of competing strategies are, from far enough back, the same kind of object, and a result proved for one can be carried to the other. Fifteen years of this work have mostly been about carrying things across that bridge.

A note on what a plain version leaves out

Every explanation above is missing the proofs, the failure cases, the ablation studies and the careful statements of what was assumed. Those are not padding. They are how a claim earns the right to be believed, and they are why the papers are written the way they are.

What a plain version can do is tell you whether the work is worth your attention, and whether the person who did it understood it well enough to say it simply. Both of those are fair tests.

The full list, with every paper and its DOI, is on the publications page.