Roee Hendel

I was most recently an Algorithm Developer at AI21 Labs.
There I worked on engineering challenges at the boundary of research, particularly in large-scale distributed training of LLMs with reinforcement learning.
Earlier at AI21 Labs, I worked on data, evaluation, and experimentation for fine-tuning (SFT), and on retrieval-augmented generation.

Overall, I have seven years of R&D experience in machine learning across deep learning, natural language processing, computer vision, and reinforcement learning, spanning both academic and industrial research.
I am a graduate of the Talpiot program, through which I completed a B.Sc. in Physics and Computer Science, and I hold an M.Sc. in Computer Science from Tel Aviv University.

I’m excited about tackling the core challenges toward achieving useful, efficient, and safe general intelligence.

selected publications

In-Context Learning Creates Task Vectors

Roee Hendel, Mor Geva, and Amir Globerson

EMNLP 2023 Findings, 2023

A language model distills the examples in its prompt into a single internal vector that represents the task itself, then answers new queries from that vector alone.

arXiv PDF
Reducing LLM Training Waste with Model-Agnostic Padding Minimization

Roee Hendel

AI21 Labs technical blog, 2026

Because it removes padding entirely outside the model, this method fits any architecture, unlike tailored approaches such as sequence packing, yet recovers almost all of their speedup. Used to train Jamba 2 efficiently.

Website
Structured RAG for Answering Aggregative Questions

Omri Koshorek, Niv Granot, Aviv Alloni, and 6 more authors

arXiv preprint arXiv:2511.08505, 2025

Standard retrieval breaks down on questions that have to aggregate over a whole corpus rather than a few passages. We pinpoint that failure and propose a first direction toward solving it.

arXiv PDF