InSTA: Towards Internet-Scale Training For Agents

Brandon Trabucco , Gunnar A. Sigurdsson , Robinson Piramuthu , Ruslan Salakhutdinov

🏛 Institutions: CMU , Amazon
📅 Date: February 10, 2025
📑 Publisher: arXiv
💻 Env: Web
🔑 Keywords: internet-scale training web agents synthetic tasks trajectory data browser automation InSTA

TLDR

InSTA studies how to scale web-agent training by turning large numbers of live websites into instruction-following browser tasks. It builds a 150K trajectory dataset with automated task generation, execution, and filtering, then trains agents that improve web navigation performance across multiple benchmarks.

Open paper arXiv Report issue