WebVoyager: Building an End-to-End Web Agent with Large Multimodal Models

TL;DR

WebVoyager is a novel multimodal web agent powered by large models that interacts with real websites to complete user tasks, establishing a new benchmark and evaluation protocol for real-world web applications.

Contribution

Introduction of WebVoyager, a large multimodal model-based web agent capable of end-to-end interaction with real websites and a new benchmark with automatic evaluation protocol.

Findings

WebVoyager achieves 59.1% task success rate on the new benchmark.

The automatic evaluation protocol has 85.3% agreement with human judgment.

WebVoyager significantly outperforms GPT-4 (All Tools) and text-only setups.

Abstract

The rapid advancement of large language models (LLMs) has led to a new era marked by the development of autonomous applications in real-world scenarios, which drives innovation in creating advanced web agents. Existing web agents typically only handle one input modality and are evaluated only in simplified web simulators or static web snapshots, greatly limiting their applicability in real-world scenarios. To bridge this gap, we introduce WebVoyager, an innovative Large Multimodal Model (LMM) powered web agent that can complete user instructions end-to-end by interacting with real-world websites. Moreover, we establish a new benchmark by compiling real-world tasks from 15 popular websites and introduce an automatic evaluation protocol leveraging multimodal understanding abilities of GPT-4V to evaluate open-ended web agents. We show that WebVoyager achieves a 59.1% task success rate on our benchmark, significantly surpassing the performance of both GPT-4 (All Tools) and the WebVoyager (text-only) setups, underscoring the exceptional capability of WebVoyager. The proposed automatic evaluation metric achieves 85.3% agreement with human judgment, indicating its effectiveness in providing reliable and accurate assessments of web agents.