Cover Image

Saved in:
Bibliographic Details
Main Authors: Andreux, Mathieu, Skuk, Breno Baldas, Benchekroun, Hamza, Biré, Emilien, Bonnet, Antoine, Bordie, Riaz, Bout, Nathan, Brunel, Matthias, Cedoz, Pierre-Louis, Chassang, Antoine, Chen, Mickaël, Constantinou, Alexandra D., d'Andigné, Antoine, de La Jonquière, Hubert, Delfosse, Aurélien, Denoyer, Ludovic, Deprez, Alexis, Derupti, Augustin, Eickenberg, Michael, Federico, Mathïs, Kantor, Charles, Koegler, Xavier, Labbé, Yann, Lee, Matthew C. H., de Kergaradec, Erwan Le Jumeau, Mahla, Amir, Manevich, Avshalom, Maret, Adrien, Masson, Charles, Maurin, Rafaël, Mena, Arturo, Modard, Philippe, Moyal, Axel, Kerbel, Axel Nguyen, Revelle, Julien, Richter, Mats L., Santos, María, Sifre, Laurent, Theillard, Maxime, Thibault, Marc, Thiry, Louis, Tronchon, Léo, Usunier, Nicolas, Wu, Tony
Format: Preprint
Published: 2025
Subjects:
Artificial Intelligence
Online Access:https://arxiv.org/abs/2506.02865
Tags: Add Tag
No Tags, Be the first to tag this record!
Table of Contents:
  • We present Surfer-H, a cost-efficient web agent that integrates Vision-Language Models (VLM) to perform user-defined tasks on the web. We pair it with Holo1, a new open-weight collection of VLMs specialized in web navigation and information extraction. Holo1 was trained on carefully curated data sources, including open-access web content, synthetic examples, and self-produced agentic data. Holo1 tops generalist User Interface (UI) benchmarks as well as our new web UI localization benchmark, WebClick. When powered by Holo1, Surfer-H achieves a 92.2% state-of-the-art performance on WebVoyager, striking a Pareto-optimal balance between accuracy and cost-efficiency. To accelerate research advancement in agentic systems, we are open-sourcing both our WebClick evaluation dataset and the Holo1 model weights.

Similar Items