Ben Finkelshtein

Passionate Research Scientist at Google Research, focusing on large-scale graph and tabular foundation models. PhD student at the University of Oxford, supervised by DeepMind Chair of AI Prof. Michael M. Bronstein and Prof. Ismail İlkan Ceylan.

For further information, here is my CV:

news

Aug 31, 2026	Started a Research Scientist full-time role at Google Research – creating the next large Graph Foundation Model.
Jan 25, 2026	Actions Speak Louder than Prompts: A Large-Scale Study of LLMs for Graph Inference is selected for an ORAL at ICLR 2026
Jan 25, 2026	Efficient Learning on Large Graphs using a Densifying Regularity Lemma is accepted to ICLR 2026
Sep 18, 2025	Equivariance Everywhere All At Once: A Recipe for Graph Foundation Models is accepted to NeurIPS 2025
Sep 15, 2025	Started an internship at Google Research – creating the first large Graph Foundation Model.

latest posts

Mar 03, 2026	Actions Speak Louder Than Prompts: Rethinking How LLMs Reason Over Graph Data - Microsoft Research
Jan 16, 2024	Graph & Geometric ML in 2024: Where We Are and What's Next (Part I - Theory & Architectures) \| Towards Data Science
Dec 06, 2023	Attention Required! \| Cloudflare

selected publications

Cooperative Graph Neural Networks

Ben Finkelshtein, Xingyue Huang, Michael Bronstein, and 1 more author

ICML, 2023

Abs HTML

A more dynamic and flexible message-passing paradigm in which each node can choose a different communication strategy.
Position: Graph Learning Will Lose Relevance Due To Poor Benchmarks

Maya Bechler-Speicher, Ben Finkelshtein, Fabrizio Frasca, and 9 more authors

ICML, 2025

Abs HTML

This paper calls towards more meaningful benchmarks, rigorous evaluation protocols, and stronger collaboration with domain experts to drive advances in graph learning research.
Equivariance Everywhere All At Once: A Recipe for Graph Foundation Models

Ben Finkelshtein, İsmail İlkan Ceylan, Michael Bronstein, and 1 more author

NeurIPS, 2025

Abs HTML

We derive the first graph foundation model that is provably universal and achieves SOTA zero-shot performance.
Actions Speak Louder than Prompts: A Large-Scale Study of LLMs for Graph Inference

Ben Finkelshtein, Silviu Cucerzan, Sujay Kumar Jauhar, and 1 more author

ICLR, 2026

Abs HTML

A principled analysis of LLM capabilities and limitations in processing graph information.