WebWorld: The Browser as a World Model for Self-Improving Web Code
v1
Paper introduction · English narration and external captions
v1

Paper: WebWorld: The Browser as a World Model for Self-Improving Web Code (v1) https://arxiv.org/html/2608.30530v1
A web repair has two jobs: establish the requested behavior and keep the behavior already verified. WebWorld makes both obligations part of its repair contract and its training-data admission rule.
WebWorld treats the browser as an executable simulator for an HTML artifact under user actions. The runtime renders pages, applies probes and records screenshots, document state, console information and action traces. A vision-language model can critique those observations, but the critique itself is not a repair certificate.
An interaction contract specifies the target predicate, replay actions, preservation set and impact scope. A typed repair skill proposes a bounded patch; filters narrow candidates before another browser execution. Acceptance then records the patch and its supporting contract evidence. These checks concern specified interactions, not every possible website action.
The same distinction governs learning. Runtime activity can help route the next attempt, but rejected or partial changes do not enter supervised training simply because the agent was busy. Only admitted transitions carry the artifact, addressed failure and evidence supporting acceptance. Previously verified predicates become later preservation obligations.
The matched experiments use 32,800 accepted transitions and keep the model family, prompt, optimizer and evaluation runner fixed within each size. The 27B HTMLBench-400 composite rises from 47.4 under raw supervision to 52.7 with WebWorld, a 5.3-score-point difference. In a separate 9B ablation, loosening certificate admission leaves a 0.4-point gain over Raw, versus 4.8 with the full gate. These are benchmark-specific score comparisons. The actionable idea is to define what changed, execute the test and check that working behavior survived.
Watch the companion English animation with external English, Simplified Chinese or bilingual captions.
Paper, exact arXiv v1: https://arxiv.org/html/2608.30530v1
#WebDevelopment #AIAgents #AIResearch
Clean English video. Choose English, Simplified Chinese or bilingual captions in the player.