WebWorld: The Browser as a World Model for Self-Improving Web Code
WebWorld makes the browser the gatekeeper for self-improving web-code training.
The paper argues that VLM repair loops fail when the same model proposes fixes and judges whether they worked. WebWorld instead has the model produce critiques, turns them into typed browser interactions, and accepts a change only if the page makes target progress while preserving previously verified behavior. Those accepted transitions become the supervision data for fine-tuning. The authors report gains over matched raw models on HTMLBench-400 and MiniAppBench-Val, with ablations attributing the lift to browser-backed certification. HF Daily Papers' note
The paper argues that VLM repair loops fail when the same model proposes fixes and judges whether they worked. WebWorld instead has the model produce critiques, turns them into typed browser interactions, and accepts a change only if the page makes target progress while preserving previously verified behavior. Those accepted transitions become the supervision data for fine-tuning. The authors report gains over matched raw models on HTMLBench-400 and MiniAppBench-Val, with ablations attributing the lift to browser-backed certification. HF Daily Papers' note
score 5