Ilmu Komputer & AI editorial
WebWorld: The Browser as a World Model for Self-Improving Web Code
The core problem
Vision-language model (VLM) driven self-improvement of web code suffers from a structural flaw: the model that proposes a repair is also the model that judges it. Under such a self-referential judge, visual plausibility becomes a poor proxy for whether the page actually works. A rendered screenshot may look correct while the underlying HTML artifact fails under real user actions, and the VLM has no incentive or mechanism to detect this gap.
The missing ingredient is a counterparty the VLM cannot fool. 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. WebWorld is presented as the interface that lets a VLM prior interact with this browser-as-world-model autonomously and decides which interactions become supervision. The central claim is that browser-backed admission, not merely more data or larger models, carries the gain in self-improving web code generation.
Innovation
Under matched training, WebWorld-27B improves Raw-27B by 5.3 points on HTMLBench-400 and 14.9 points on MiniAppBench-Val. The resulting system reaches the level of strong frontier systems such as Kimi-K2.6 and GPT-5.4 on interactive HTML generation. These gains are achieved with the same training budget and model scale as the baseline, isolating the effect of the browser-backed admission mechanism.
Equal-size ablations further show that browser-backed admission carries the gain. Without the acceptance certificate, the matched 9B lift nearly disappears. This indicates that the improvement is not merely a function of additional interaction rounds or model capacity, but specifically of the certified quality ratchet that filters transitions before they enter the SFT export. The contrast between the 27B and 9B ablation results underscores that the certificate mechanism is the load-bearing component of the method.
Why it matters
The results support a structural rather than scale-centric explanation for self-improvement in web code. When the same model proposes and judges repairs, visual plausibility under that judge is a weak proxy for functional correctness. By introducing the browser as a deterministic, executable counterparty, WebWorld converts the self-improvement loop into a sequence of verifiable transitions. The acceptance certificate enforces both target progress and preservation of previously verified capabilities, which prevents the regression that typically erodes gains in iterative self-repair.
The quality ratchet acts as a monotonic filter: only certified transitions reach the SFT export, so the training distribution is composed exclusively of behavior that has been validated under executable user actions. This design choice explains why the matched 9B ablation loses nearly all lift without the certificate, and why WebWorld-27B can approach frontier systems on interactive HTML generation. The broader implication is that world models for code need not be learned from scratch; existing deterministic simulators, such as the browser for web artifacts, can serve as the counterparty that grounds self-improvement. Future work may extend this pattern to other executable domains where a deterministic simulator can issue acceptance certificates.
Who should read this
Opening member content…