About
Hi! I am researching a lot of cool AI stuff at the moment. My current focus is how models can improve themselves: adapting a model to fit the task in front of it. I study whether a dedicated model can actually understand the high-dimensional space of another model's weights and adapt and improve the model in real time.
Before this, I was at IBM Research, where I worked on foundation models. I started in drug discovery, worked on foundation models for cross-modal learning (MAMMAL) and on diffusion models for structure prediction and design, where we pre-trained our foundation models on billions of samples. I then joined the Granite Vision team, which builds the vision-language model of the IBM Granite family. A recurring theme was getting different modalities to work together inside one model.
I did my MSc at Tel Aviv University, advised by Prof. Mahmood Sharif, on adversarial machine learning in computer vision and the robustness of face recognition (StyleAT, BMVC 2026).
Research
-
Now
How models can improve themselves
Adapting a model to the task in front of it. My current step: a meta-model reads a task description or a few examples and produces a weight update for a base model in one forward pass, with no gradient steps. I'm interested in what such a map can learn, why it should exist at all, and where it breaks.
-
IBM Research, Granite Vision team
Vision-language models
I worked on Granite Vision, the vision-language model of the IBM Granite family, built to understand documents, charts and tables alongside text.
-
IBM Research, drug discovery
Multimodal integration
Getting models of different modalities to work as one. I worked on aligning biomedical encoders to large language models so they can reason over proteins and molecules alongside text, and on fusing protein language models, LLMs and 3D structural models into unified protein representations.
-
IBM Research, drug discovery
Foundation models for drug discovery
Biomedical foundation models across proteins, small molecules and genes: a multi-aligned foundation model that handles all three in one architecture (MAMMAL), latent-variable transformers for controllable molecule generation (STAR-VAE), and diffusion models for biomolecular structure prediction and design.
-
MSc, Tel Aviv University
Adversarial machine learning
Robustness of face recognition to semantic attacks, where the adversary changes attributes such as age or pose instead of adding pixel noise. We built BoundStyle, an attack in StyleGAN's latent space that runs about 9.5 times faster than earlier attacks, and StyleAT, an adversarial training scheme that uses it to defend against attacks it has not seen.
Publications
-
StyleAT: Defending Face Recognition Against Semantic Attacks
-
MAMMAL: Molecular Aligned Multi-Modal Architecture and Language for Biomedical Discovery
-
STAR-VAE: A Scalable Latent-Variable Transformer for Controllable Molecular Generation
-
BioVERSE: Representation Alignment of Biomedical Modalities to LLMs for Multi-Modal Reasoning
-
Cross-Modal Fusion of Protein Language Models, Large Language Models, and 3D Structural Models for Unified Protein Representations