Don’t Retrain, Remix: Adaptation of Web Agents using Mixture Hypernetworks

Humans compose and transfer skills with little effort. Someone who has booked a flight on American Airlines can reuse the same search, filter, and checkout flow on an unfamiliar credit card travel portal. When a familiar site is redesigned, regular users need a few visits to adjust, not to relearn the task. Fine-tuned web agents struggle with both kinds of change. Our goal is to help web agents reuse and combine learned skills to generalize to unseen domains and new versions of familiar interfaces.