The BrowserGym Ecosystem for Web Agent Research

Source: arXiv Authors: Thibault Le Sellier De Chezelles, Maxime Gasse, Alexandre Drouin, Massimo Caccia, Léo Boisvert, Megh Thakkar, Tom Marty, Rim Assouel, et al. Date: 2024-12-06 Primary category: cs.LG All categories: cs.LG, cs.AI, cs.SE

Abstract

BrowserGym turns web-agent evaluation into a unified gym-like substrate with standard observations and actions across multiple benchmarks. It is one of the clearest examples of an actual “gym” for agent harnesses rather than merely a benchmark paper with a leaderboard attached.