Imubit's research starts with the mathematics of why neural networks work, and when they can be trusted to act. That foundation sits under every Deep Learning Process Control® controller running a unit today, inside the envelope its engineers set.
Co-founder and CTO, ImubitAssociate Professor of Computer Science and AI, Tel Aviv University
Nadav leads research that runs from the foundations of deep learning to its use in the physical world. His academic group at Tel Aviv University develops mathematical theories of neural networks and the experiments that test them. At Imubit, his teams build the models that make decisions on running plants.
Most AI is built to predict. Control is harder: the model's output moves a real unit, and a wrong move costs product, energy or safety margin. Nadav's ERC-funded research develops a theory of neural networks for control, to understand mathematically when a learned controller behaves well and why, rather than only testing it after the fact.
How Control works
His group showed that an agent that learned to act safely on one set of tasks does not, by that fact, act safely on new ones: safe execution transfers worse than execution itself. That is why a DLPC® is built and validated for each unit, and runs inside the envelope that unit's engineers set.
Read the spotlight
A Deep Learning Controller is trained by reinforcement learning across millions of simulations of the unit, then runs as a deterministic controller: the same conditions get the same move, every time. Closed loop never runs through a language model.
What Closed-Loop Physical AI meansNadav's work, written up for the people who run plants. New spotlights are added as they publish.
Nadav Cohen's group at Tel Aviv University proved that safe execution of a task does not transfer the way execution does, so a general-purpose agent that was safe elsewhere is not, by that fact, safe on your unit.
Imubit's co-founder and CTO Nadav Cohen and his Tel Aviv University students received the Israel AI Safety Research Prize for work published at NeurIPS 2025, selected as a Spotlight paper.
Correlation is enough to predict. Acting on a running unit needs cause and effect. What the difference means for AI in process control.
Imubit's co-founder and CTO Nadav Cohen on TheMarker's podcast: why losing control of the internet is a nearer AI risk than doomsday scenarios, and why the physical world moves more carefully.
One to two hours for your team to prepare and upload the data package
Results within a few working days of approval
No workshops
Send your data as it is. Gaps, outages and noisy tags are normal, and our engineers screen and clean them in the platform. If anything limits a use case, we tell you.
