WebMix lets web agents adapt to new sites without retraining

Researchers introduced WebMix, a mixture hypernetwork that lets web‑automation agents handle unseen website layouts without gradient updates. The system keeps a bank of low‑rank adapters trained on known sites. When faced with a new site, a scout agent explores it, creates a short manual, and a router predicts how to mix the adapters into a new one. The adaptation needs only the scouting run and a forward pass. Tests with Qwen and Llama models on TimeWarp and WebArena‑Lite‑v2 showed WebMix matching or beating in‑context learning and LoRA baselines.