Abstract
WebWorld uses a browser as an executable world model to certify web code repairs via interaction contracts, enabling autonomous VLM self-improvement with verified supervision.
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.
Community
🌐 This paper introduces WebWorld, a system that fundamentally resolves the structural flaw of VLM-driven web-code self-improvement—where the same model acts as both proposer and judge—by treating the browser as a deterministic, executable world model; the core mechanism strictly confines the VLM to hypothesis generation, mandating that every candidate repair must be re-executed by the browser and issued an acceptance certificate (requiring both target progress and preservation of all prior capabilities) before entering the SFT training pool, thereby shifting supervision from visual plausibility to verifiable behavioral evidence. Experiments show that this mechanism boosts a 27B model by 5.3 points on HTMLBench-400 and 14.9 points on MiniAppBench-Val, reaching frontier performance on par with Kimi-K2.6 and GPT-5.4, while ablation studies reveal that removing the browser certificate collapses the gain to merely 0.4 points—strongly demonstrating that only world-model-backed certified transitions can provide genuinely reliable supervision for self-improving web code.
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
- BrowserForge: Scaling Web Episode via Parallel Browser Sandboxes (2026)
- ASIL: Replacing Screenshot-and-Click with Structured State and Semantic Actions (2026)
- BackendForge: Benchmarking Agentic End-to-End Code Generation with Backend Services (2026)
- UI2App: Benchmarking Visual Interaction Inference in Executable Web Application Generation (2026)
- BTS-AgentBench: A Deterministic, Replayable Pipeline from Read-Only Telemetry Logs to Agent Benchmarks (2026)
- AutoCedar: An Agentic Framework for Verifier-Guided Access Control Policy Synthesis (2026)
- Interaction Scaling: Grounding the Third Axis of Test-Time Compute (2026)
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend
Get this paper in your agent:
hf papers read 2608.30530 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper