Web World Models
Authors: Jichen Feng, Yifan Zhang, Chenggong Zhang, Yifu Lu, Shilong Liu, Mengdi Wang
Paper: https://arxiv.org/abs/2512.23676
Code: https://princeton-ai2-lab.github.io/Web-World-Models/
TL;DR
WHAT was done? The authors introduce the Web World Model (WWM), a hybrid architectural paradigm that decouples environmental state into two layers: a deterministic “Physics” layer defined by standard web code (TypeScript/JSON), and a probabilistic “Imagination” layer synthesized by Large Language Models (LLMs). Through a suite of implementations—ranging from infinite procedural galaxies to card games and cellular automata—they demonstrate how to build consistent, effectively infinite environments without relying on massive static databases or hallucination-prone end-to-end generative models.
WHY it matters? This work addresses the “missing middle” between rigid, finite web applications and uncontrollable, fully generative world models (like World Models). By treating the modern web stack as a neuro-symbolic substrate, it offers a practical blueprint for deploying long-running, hallucination-free agents that can explore infinite state spaces while retaining the structural guarantees required for game mechanics and software reliability.
Details
The Reliability-Scale Trade-off
The central tension in modern agentic environment design lies between consistency and scale. Traditional web frameworks offer robust state management and determinism via databases (e.g., SQL), but the world size is strictly bounded by stored data. Conversely, fully generative approaches—such as Generative Agents (see review) or learned simulators like WebDreamer—promise infinite context but suffer from “dream-like” instability, where object permanence dissolves and physical rules drift. The authors identify that current approaches force a choice: a small, reliable world or an infinite, unreliable one. The Web World Model (WWM) proposes that we do not need to choose if we architecturally separate the rules of the world from its content.
