WebWorld-27B lifts HTMLBench 5.3 points with browser verifier

A new arxiv paper hands the job of grading a VLM's web-code repairs to the browser itself, and keeps only the browser's verdict as training data.

WebWorld-27B improves its raw base by 5.3 points on HTMLBench-400 and 14.9 points on MiniAppBench-Val, reaching the level of frontier systems the paper names as Kimi-K2.6 and GPT-5.4 for interactive HTML generation.

The paper describes the loop: '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.' Only certified transitions become training data. In the paper's framing, the browser is 'a counterparty the VLM cannot fool', described as a deterministic, executable simulator of how an HTML artifact behaves under user actions.

MiniAppBench itself, per its GitHub repo, holds 500 tasks across six domains: Science, Games, Tools, Visualization, Humanities and Lifestyle.

It lands the same week that World Labs pushed its Atlas omni world model into early access, another release built around treating an environment as ground truth for a model.