I’m a first-year PhD candidate in the Schwartslab of Huji-NLP at the Hebrew University.
My research focuses on leveraging compression-optimized tokenizers to boost the efficiency and interpretability of language models while also exploring multimodality to build more robust and scalable systems.
When I’m not delving into cutting-edge NLP challenges, I’m a passionate Tel Aviv fan, a devoted cat lover, a sea enthusiast, and arguably the most enthusiastic reader around.
Education
Hebrew University of Jerusalem
MSc. in Computer Science
Advisor: Prof. Roy Schwartz
Thesis: Detokenization in LLMs
The Open University
BSc in Computer Science
Tel Aviv University
BA. in Economics
Haifa University
BA. in General Studies
Publications
preprint
Follow the Flow: On Information Flow Across Textual Tokens in Text-to-Image Models
Guy Kaplan*, Michael Toker*, Yuval Reif, Yonatan Belinkov, Roy Schwartz
The study shows that only a few tokens per word capture its core meaning, leaving many redundant tokens that can impair image generation. Removing these redundant tokens improves generation quality, while patching leaked token representations with their isolated versions effectively mitigates semantic leakage.
Experience
Chief Scientist Officer — BRIGHT
Developing model based system for assisting forensics odontologists in identifying human remains
Data Science & SWE — Microsoft
Developed novel algorithms for risk score user assessment over Microsoft Defender platform
SWE Intern — Yahoo!
Worked on improving personal recommendations for Yahoo! advertising platform
Portfolio
Social Media based Stock Prediction
Developed a model to predict stock prices based on social media sentiment analysis