The Universal Normal Embedding
A conceptual framework proposing a shared Gaussian latent space underlying both vision encoders and generative models.
PhD Student · Senior Researcher
Roy Betser
I am a researcher working at the intersection of machine learning theory, representation learning, and trustworthy AI. I received my BSc. and MSc. from the ECE faculty at the Technion, where I continue my PhD studies. As a PhD candidate in the Technion, under the supervision of Prof. Guy Gilboa, I study the geometric and probabilistic structure of learned representations. My work spans theory, conceptual frameworks, and practical methods, aiming to uncover underlying structure in foundation vision representations and leverage it for a range of applications. In parallel, as a Senior Researcher at Fujitsu Research of Europe, I work on AI security, safety, and trust. This work focuses on detection and evaluation methods to ensure safety, reliability, and control in the rapidly evolving landscape of LLM-based agentic systems.
Selected Publications
For a full list, see scholar profile (link above).
Training-free Detection of Generated Videos via Spatial-Temporal Likelihoods
Detecting generated videos via combined spatial and temporal likelihood estimation in pretrained vision representations.
InfoNCE Induces Gaussian Distribution
A theoretical analysis showing how Gaussian structure emerges in learned representations under contrastive objectives.
Training-Free Policy Violation Detection via Activation-Space Whitening in LLMs
LLM policy violation detection through activation-space whitening of hidden representations.
General and Domain-Specific Zero-shot Detection of Generated Images via Conditional Likelihood
A domain-adaptive approach for detecting generated images via conditional likelihood estimation in CLIP embedding space.
CAIR : Counterfactual-based Agent Influence Ranker for Agentic AI Workflows
LLM policy violation detection through activation-space whitening of hidden representations.
Whitened CLIP as a Likelihood Surrogate of Images and Captions
A whitening-based Gaussian modeling of CLIP embeddings that yields a closed-form likelihood surrogate for images and captions.