Skip to content

CornellDataScience/world-wide-web-2

Repository files navigation

World Wide Web

A web navigation AI agent trained with reinforcement learning, using a learned world model as a simulator. The agent learns to complete tasks on real websites — like searching for information, filling out forms, and navigating multi-page workflows — without needing a live browser for every training step.


Overview

Training web agents the traditional way is slow and fragile. Every action requires waiting on a real browser to load a page, and one network hiccup can crash a training run. This project takes a different approach:

  1. Learn a world model — a smaller LLM that predicts how a webpage changes when an action is taken
  2. Train an RL agent against the world model — the agent "dreams" simulated rollouts at GPU speed instead of waiting on browsers

The result is a navigation policy that can be trained orders of magnitude faster than browser-in-the-loop methods.

About

World Wide Web 2 teaches an AI to browse the web by training it inside a learned simulator instead of a real browser. This makes training much faster and lets the agent learn from far more practice without needing live websites.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors