The authors present WebWorld, an interface that allows a Vision-Language Model (VLM) to interact autonomously with the browser as a deterministic world model for web code. This approach addresses the structural flaw in VLM-driven self-improvement where the model judges its own repairs by introducing a counterparty that cannot be fooled.

  • The VLM emits critiques compiled into interaction contracts, which the browser re-executes to issue acceptance certificates based on target progress and capability preservation.
  • Certified transitions accumulate as a quality ratchet used for Supervised Fine-Tuning (SFT) export.
  • WebWorld-27B improves Raw-27B by 5.3 points on HTMLBench-400 and 14.9 points on MiniAppBench-Val.
  • The system reaches performance levels comparable to Kimi-K2.6 and GPT-5.4 on interactive HTML generation.
  • Ablations confirm that browser-backed admission is responsible for the gains, as removing the certificate causes the lift to nearly disappear.

This method provides a rigorous verification loop that ensures web code actually works under user actions rather than just appearing visually plausible.