Beyond Pixel Histories: World Models with Persistent 3D State

A research team has introduced PERSIST, a world model that simulates the evolution of a latent 3D scene to generate coherent, evolving 3D worlds with persistent spatial memory and consistent geometry, according to a paper posted on arXiv [1]. The framework, detailed in a submission last revised on 3 June 2026, departs from existing interactive world models that rely on pixel-based histories and limited temporal context windows [1][2]. Instead, PERSIST decomposes world simulation into three coupled components: a world-frame model that predicts how a learned representation of a 3D scene evolves over time, a camera model that tracks the agent’s viewpoint, and a world-to-pixel generation module that produces observations via differentiable projection and a learned rendering function [3]. “Rather than treating pixels as the primary carrier of memory, PERSIST models the structure and dynamics of the environment in a latent 3D space,” the authors write [4]. The camera acts as a query mechanism that extracts the subset of 3D information relevant to guide the generation of the current frame [4]. This formulation enables long-horizon rollouts with fixed-cost memory and enforces geometric consistency by construction, while remaining compatible with modern diffusion and flow-based training objectives [5]. World models, which build internal representations of an environment and predict how it changes over time in response to actions, have been a focus of AI research since the 1990s and now power applications in robotics, autonomous driving, and interactive video generation [6]. The PERSIST framework was evaluated in a complex 3D environment and showed substantially improved long-horizon generation quality compared to rolling-window and memory-retrieval baselines, with stronger temporal stability, improved spatial memory when revisiting previously observed regions, and markedly better 3D consistency overall [3][5]. The authors also demonstrate novel capabilities including synthesising diverse 3D environments from a single image, explicit 3D scene initialisation, mid-episode scene edits, and the emergence of off-screen dynamic processes that continue to evolve even when not directly observed [1][4]. The paper was first submitted on 3 March 2026 by Samuel Garcin and colleagues, with an initial submission size of 19,623 KB; the revised version weighed 15,543 KB [1].