Training Object Permanence in World Models

2026-09-23

AI Summary

This study addresses the lack of physical cognitive priors, such as object permanence and entityness, in existing video generation models. We propose WROP, a data infrastructure that leverages Blender's procedural generation to randomize parameters while preserving underlying cognitive structures. Furthermore, we introduce the first cognitive science-inspired dataset and evaluation benchmark comprising 150 tasks, establishing an object permanence training paradigm specifically designed for world models. Built upon a large-scale synthetic data pipeline and an AWS Trainium2-native PyTorch training stack, this project releases a 1.5M-sample corpus and trains the PWM-WROP model. In blind Elo ranking evaluations, the proposed model achieves first place among video continuation models and third place overall.

Abstract

Object permanence and solidity are hallmarks of human cognitive priors. Recent studies show that video generation models, a paradigmatic class of current world models, have begun to show emerged reasoning abilities, making them ideal candidates for building human-like physical intelligence. Do video models have emerged object permanence in them? If not, could we train them with a core-cognition inspired dataset? We introduce WROP (World Reasoning with Object Permanence), a data infrastructure of 150 hand-designed cognitive science inspired tasks, divided into six cognitive categories. We build Blender generators that randomize speed, lighting, camera angle, and other nuisance parameters while preserving each task's cognitive structure, yielding 10,000+ samples per task. We release a 1.5M-sample training corpus and a 300-question exam. On this exam we evaluate 14 video models: 3 reference-to-video, 7 edit, and 4 continuation, among which PWM-WROP, our 16B world model. In a blind pairwise Elo study, PWM-WROP ranks first among continuation models and third overall, behind only a statistical tie between two reference-to-video models. We release the data, exam, model answers, scores, weights, and PWM, our native-PyTorch training stack on AWS Trainium2.

Problem

• Object Permanence
• World Models
• Video Generation Models
• Physical Intelligence
• Core Cognition

Innovation

• Object Permanence
• World Models
• Video Generation
• Cognitive Science
• Data Infrastructure