arXiv:2609.10540 · cs.CV · 2026
Programmable World Model
Zheng-Hui Huang, Guixu Lin, Jiacheng Lin, Yi-Chuan Huang, Ruihan Yu, Muyao Niu, Siqi Yang, Yu-Lun Liu, Yung-Yu Chuang, Kaipeng Zhang, Zhixiang Wang
17 citations · 3 influential citations— Semantic Scholar
TL;DR · Semantic Scholar
The Programmable World Model is introduced, a framework that decouples world-state evolution from visual observation generation and demonstrates the effectiveness of separating explicit state evolution from generative rendering for building persistent, programmable worlds.
Saved summary
AI-generated while a reader studied this paper. Check the original source for accuracy.
Problem: Interactive video world models lack an explicit, persistent global state that can be updated independently of the current viewpoint. Consequently, they cannot easily handle off-screen entities, keep track of nonvisual state facts, or allow users to program explicit rules governing world evolution. This reliance on implicit state representations in generative contexts often leads to accumulated update errors, disappearing characters, and physical inconsistencies.
Method: The Programmable World Model tracks explicit world-space updates using state-augmented 3D oriented bounding boxes (OBBs) inside a lightweight engine. A deterministic state compiler projects these OBBs under a target camera to generate pixel-aligned identity, semantic, and direction maps. These maps act as spatial controls for a generative renderer built on LingBot-World-v1 with a Structured Spatial ControlNet, leveraging multi-scale temporal history and a geometry-aligned spatial memory from AlayaWorld.
Results: On the CombatStateBench benchmark, the proposed method achieves a Count Accuracy of 94.00\% and a State Accuracy of 98.00\%, significantly outperforming baselines like LingBot-World-V2 and YUME. It also obtains superior performance across VBench quality metrics, reaching 67.62\% for Imaging Quality, 94.74\% for , for , and for .