Search papers, labs, and topics across Lattice.
This paper introduces WebWorld, a novel framework that leverages a browser as a deterministic world model to enhance the self-improvement of web code through visual language models (VLMs). By allowing the VLM to interact with the browser to validate code changes, WebWorld addresses the inherent flaw of self-judgment in VLMs, ensuring that only functionally sound modifications are accepted. The results show that WebWorld-27B significantly outperforms Raw-27B and matches the performance of leading systems in interactive HTML generation, demonstrating the effectiveness of browser-backed validation in this context.
WebWorld reveals that integrating a browser as a world model can drastically enhance the reliability of VLM-driven web code improvements, achieving performance on par with top-tier systems.
VLM-driven self-improvement of web code has a structural flaw: the model that proposes the repair is the model that judges it, and visual plausibility under that judge is a poor proxy for whether the page actually works. What the loop is missing is a counterparty the VLM cannot fool, and the browser already is that counterparty: a deterministic, executable simulator of how an HTML artifact behaves under user actions, and in everything but name a world model for web code. We present WebWorld, the interface that lets a VLM prior interact with this browser-as-world-model autonomously and decides which interactions become supervision. Each round, the VLM emits a critique that the planner compiles into a typed interaction contract; the browser re-executes the candidate and issues an acceptance certificate only when both target progress and preservation of every previously verified capability hold; certified transitions accumulate as a quality ratchet that is the only thing the SFT export ever sees. Under matched training, WebWorld-27B improves Raw-27B by 5.3 points on HTMLBench-400 and 14.9 points on MiniAppBench-Val, and reaches the level of strong frontier systems such as Kimi-K2.6 and GPT-5.4 on interactive HTML generation. Equal-size ablations show that browser-backed admission carries the gain: without the certificate, the matched 9B lift nearly disappears.
