I am a researcher at MSL FAIR and affiliated with Tel Aviv University, working on the theory of machine learning. I received my PhD in theoretical fundamental physics at Tel Aviv University and then worked as a postdoctoral researcher at EPFL on physics of AI. My research focus centers on how the structure of data interacts with the learning mechanism to shape what models learn, and how that understanding can inform their emergent behaviors, capabilities, alignment and safety at scale.

I am requesting $2,800 to attend NeurIPS Paris, where I have been invited to speak at the Principles of Generative Modeling workshop on December 12, 2026. I will discuss inference scaling and reasoning models, and also plan to present work with colleagues from Stanford on efficient pass@k sampling for LLM reasoning. The workshop has no budget for my travel or accommodation.

I believe scientific progress depends on researchers exchanging ideas and subjecting them to criticism, and that modern machine learning is an emerging science, lacking much of the methodological honesty of basic sciences (for instance, the meaning of "theory" is subjective, and there is no agreed upon level of "error bars" or reproducibility). This matters particularly in a field developing as quickly as AI, with as big of an impact on the real world. Workshops offer a way for researchers across institutions the opportunity to discuss, find weaknesses in their arguments and create new collaborations. That is the purpose of this trip.

The research direction I want to advance is a deeper understanding of why models behave as they do from a physical perspective, using the simplest models that capture the interesting phenomena. Empirical evaluations can only tell us how a model performs under certain conditions, while understanding the mechanisms underlying that behavior can help us reason about when those results will persist and what are their failure modes, which is particularly crucial for safety and alignment. My work spans the relationship between data structure, learning and inference, including how reasoning models use additional computation, particular with multi-agent swarms and recursive self-improvement, all of which are subjects that will be discussed in the workshop and at NeurIPS in general. I believe this understanding is necessary for building aligned and safe systems that can benefit society.

I employ tools from theoretical physics to study learning and generative models. I have already pursued this direction in multiple papers, and work on these problems every day. The invitation gives me a concrete opportunity to discuss my work with researchers studying complementary aspects of generative modeling. I want to use those discussions to find new interesting questions and build collaborations around them.
I am particularly interested in collaborating across institutions to tackle the question: how does the underlying structure of data affect alignment in generative models? I believe this is an open question, and my experience shows that the answer can depend on many factors (whether models are just pretrained, RLed, or RLHFed). Building simple models of data that can show where misalignment can emerge and how, and then testing these on real settings would be a very powerful result to come out of this visit.

The planned trip is December 11–15, including four nights in Paris. My current budget is $850 for a return flight, $1,380 for accommodation, a $240 registration fee, $80 for transport, and $250 for food and incidentals, which is approximately $2,800 in total.

Without support, I would pay for the trip personally. This grant would support my participation in an existing scientific program and an opportunity to develop the research relationships surrounding my work.

Google Scholar
PriGM workshop