flâneur

WebArena-Infinity: Generating Browser Environments with Verifiable Tasks at Scale

webarena.dev · 3,618 words · saved by 1 readers

Realistic web environments with verifiable tasks are essential for training and evaluating general-purpose browser-use agents. However, constructing such environments remains prohibitively labor-intensive, requiring manual effort across sandbox creation, data curation, task design, and programmatic verifier authoring1234. The required expertise, often including significant programming effort, further limits scalability, confining large-scale environment construction largely to well-resourced organizations. We aim to automate what has traditionally been a manual, expert-driven process. In this work, we propose an approach for automatically generating high-authenticity, high-complexity, and RL-ready environments, along with verifiable tasks, from static artifacts such as recorded workflows and user manuals using a multi-agent system. The resulting pipeline is both time- and cost-efficient: each environment can be generated within ten hours at a cost under $100, and the process is highly

GitHub Environment Hub Dataset browser-use agents navigating generated environments across diverse application domains. Realistic web environments with verifiable tasks are essential for training and evaluating general-purpose browser-use agents. However, constructing such environments remains prohibitively labor-intensive, requiring manual effort across sandbox creation, data curation, task design, and programmatic verifier authoring1234. The required expertise, often including significant programming effort, further limits scalability, confining large-scale environment construction…

saved by

related reading