Liane Galanti

Hello! I’m a first-year CS Ph.D. student at Princeton University, working with Elad Hazan, Tri Dao, and Dhruv Shah. I’m grateful to be supported by the Francis Robbins Upton Fellowship. I’m currently a Research Intern on RL & Post-Training Systems (Turbo team) at Together AI. Previously, I was a Research Scientist at IBM Research, and I received my M.Sc. in Computer Science from Tel Aviv University, where I worked with Lior Wolf.

I’m interested in questions related to AI for robotics — building robot policies that are robust, data-efficient, and fast.

I’d love to chat or work together! Reach me at galanti [at] princeton [dot] edu.

News

2026I’ll be joining Together AI — so excited! 🎉
2026Our paper on selective state-space layers was accepted to AISTATS 2026!
2025Honored to receive the Francis Robbins Upton Fellowship. 🙏
2025Excited to start my CS PhD at Princeton!
2023Two papers accepted at NeurIPS 2023 and TMLR!

Selected Publications

On the Expressivity of Selective State-Space Layers: A Multivariate Polynomial Approach

Edo Cohen-Karlik*, Itamar Zimerman*, Liane Galanti*, Ido Atad, Amir Globerson, Lior Wolf

AISTATS, 2026

paper

We show that selective state-space layers are more expressive than linear transformers, enabling richer sequence representations while preserving generalization and computational efficiency.

Norm-based Generalization Bounds for Sparse Neural Networks

Tomer Galanti, Mengjia Xu, Liane Galanti, Tomaso Poggio

NeurIPS, 2023

paper code

Derived relatively tight generalization bounds for sparse neural networks (e.g., convolutional networks), shedding light on the pivotal influence of the network’s sparsity on its ability to generalize. We empirically showed that our bound is significantly tighter than alternative generalization bounds for convolutional networks from the literature.

Comparative Generalization Bounds for Deep Neural Networks

Tomer Galanti, Liane Galanti, Ido Ben-Shaul

TMLR, 2023

paper code

Observed that neural networks trained for classification exhibit a property in which Neural Collapse propagates backward from the penultimate layer to several preceding layers. We introduced the notion of “effective depth,” which captures the number of bottom layers in a network that do not experience neural collapse, and showed empirically that a model’s effective depth is independent of its actual depth and adapts to the complexity of the data.

From Grounding to Planning: Benchmarking Bottlenecks in Web Agents with Mind2Web

Segev Shlomov, Ben Wiesel, Aviad Sela, Ido Levy, Liane Galanti, Roy Abitbol

ECAI, 2025

paper

We explored whether action planning or element grounding capabilities are more crucial for executing action-based tasks from natural-language queries (e.g., booking a flight, searching for content on Google). Using the Mind2Web benchmark, we tested various agents, isolating and assessing the contributions of planning and grounding capabilities to performance.