Persistent Computational State: A Session-Centric Runtime for Generative World Models

A new paper on arXiv challenges the prevailing view that current video world models fail as simulators due to architectural or training deficiencies. The authors demonstrate that the real issue lies in request-centric serving, which discards runtime state. By snapshotting and restoring the runtime state—including observations, RNG state, memory banks, or windowed KV context—they achieve byte-identical continuations after excursions. This shows that the capability was never missing; it was simply discarded by the serving infrastructure. The paper defines Persistent Computational State as a session-centric runtime that preserves the non-recomputable kernel of world-model state, enabling genuine backtracking and forking. The work has implications for generative world models used in planning and simulation.