Israel · 2026

TLV AI 50

The 50 Israelis Expanding the Frontier of AI

Intro

Israel has always punched above its weight in technology, and artificial intelligence is no exception. The TLV AI 50 maps the people carrying that weight right now: researchers whose ideas shape how modern models think, see, and reason, and founders turning those ideas into the chips, companies, and products the world runs on.

They work across Tel Aviv, Jerusalem, Haifa, Boston and San Francisco. At Weizmann, Technion, OpenAI, DeepMind, Meta, NVIDIA, and in stealth mode. What unites them is a distinctly Israeli mix of academic depth, engineering audacity, and impatience with the impossible. This is not a ranking - it is a snapshot of a community expanding the frontier of AI, and an invitation to watch where it pushes next.

This year, we also felt it appropriate to highlight a few noteworthy papers from the past year, authored or co-authored by up-and-coming Israeli researchers - find them below the list.

Trends

Up until this year, Israel remained mostly out of the game when it came to innovation on the model layer. This past year though has seen the emergence of several Israeli neolabs. It will be interesting to track their progress and view the unique form that Blue and White labs take.

Without a doubt though, the two areas Israel seems to be firmly at the frontier of are: AI Security and Vision.

A new cohort of AI natives are defining how AI systems are hardened, verified and deemed safe for the public. These don't look like the cyber companies of yore, that sold top down to CISOs but rather are deeply technical, research-led companies that are setting the standards for how the largest companies in the world deploy AI safely.

When it comes to vision, Israel hosts some of the most often cited academic labs. Additionally, Israelis hold key roles in most of the major labs' vision projects. Excitingly, a few startups have begun to harness the depth of talent that exists in the country and help expand the horizons of physical AI.

50People
30Researchers
20Founders
7Papers
See the 2025 list

Noteworthy Papers

Recent papers from the past year with at least one Israeli author.
arXiv · 2026

LTX-2: Efficient Joint Audio-Visual Foundation Model

Yoav HaCohen, Zeev Farbman et al. (Lightricks)
Open-source foundation model that generates video and synchronized audio (speech, foley, ambience) in one pass - an asymmetric dual-stream transformer (14B video + 5B audio) coupled by bidirectional cross-attention. Matches proprietary systems at a fraction of the compute and inference time, with weights and code publicly released.
Read paper ↗
ICLR · 2026

From Tokens to Thoughts

Chen Shani, Dan Jurafsky, Yann LeCun, Ravid Shwartz-Ziv et al.
Compares how humans and 40+ LLMs organize concepts through an Information Bottleneck lens: LLMs match human category boundaries but over-compress and lose fine-grained nuance. A crisp account of where LLM thinking diverges from human cognition.
Read paper ↗
PNAS

BetaDescribe

Yonatan Belinkov's Technion group et al.
An LLM that generates rich functional descriptions of proteins directly from sequence, including proteins with no annotated homologs. Interpretability-flavored AI-for-science hitting a real bioinformatics bottleneck.
Read paper ↗
ICML · 2026

Induction Meets Biology

Mechanistically dissects how protein language models detect exact and approximate sequence repeats: induction heads plus biology-specialized neurons. A template for interpreting evolutionary reasoning in models.
Read paper ↗
arXiv · 2026

The Dual Mechanisms of Spatial Variable Binding in VLMs

VLMs bind objects to spatial relations via two parallel mechanisms; the dominant signal sits in the vision encoder, spread globally across visual tokens. Amplifying it improves spatial binding - redirects VLM interpretability toward the vision side.
Read paper ↗
arXiv · 2026

LLMs Generate Harmful Responses Using a Distinct Mechanism, Shared Across Harm Types

Identifies a sparse set of parameters driving harmful compliance; pruning them cuts harmful outputs with little capability loss, transferring across harm categories. Points to principled parameter-level safety interventions over behavioral guardrails.
Read paper ↗
Bau Lab · Northeastern

Agents of Chaos

Two-week red-teaming of autonomous LLM agents with email, Discord, filesystems and shell access. Documents agents leaking info, spoofing identities, and reporting success when system state contradicted them - evidence that agentic deployment fails in ways static evals don't catch.
Read paper ↗