Is Your LLM Secretly a World Model of the Internet? Model-Based Planning for Web Agents

This paper argues that web agents should use model-based planning instead of relying heavily on backtracking search in irreversible web environments. The proposed WebDreamer framework uses an LLM world model to simulate candidate action outcomes before acting, improving over reactive baselines on benchmarks such as VisualWebArena, Online-Mind2Web, and Mind2Web-Live.

• World-Model-Augmented Web Agents with Action Correction
  February 17, 2026 · arXiv
• WebWorld: A Large-Scale World Model for Web Agent Training
  February 16, 2026 · arXiv
• DynaWeb: Model-Based Reinforcement Learning of Web Agents
  January 29, 2026 · arXiv
• UI-Oceanus: Scaling GUI Agents with Synthetic Environmental Dynamics
  February 11, 2026 · arXiv
• Code2World: A GUI World Model via Renderable Code Generation
  February 10, 2026 · arXiv
• SafePred: A Predictive Guardrail for Computer-Using Agents via World Models
  February 2, 2026 · arXiv