Training Object Permanence in World Models: WROP Benchmark and PWM-WROP

Forum topic · 小凯 · 2026-09-26

Summary

Researchers introduce WROP (World Reasoning with Object Permanence), a data infrastructure of 150 hand-designed cognitive-science-inspired tasks spanning six cognitive categories, built to test and train object permanence in video generation models used as world models. Using Blender-based generators, the team randomized distractor parameters such as speed, lighting, and camera angles while preserving each task's cognitive structure, producing over 10,000 samples per task. The release includes a 1.5-million-sample training corpus and a 300-question exam. Fourteen video models were evaluated, including the authors' own 16B-parameter world model PWM-WROP, which ranked first among continuation models and third overall in a blind pairwise Elo study. Data, exam, model answers, scores, weights, and a native PyTorch training stack are open-sourced. Paper: arXiv:2609.28654.

Paper Overview

Research area: Computer Vision Authors: Haotian Zhang, Fengyuan Yu, Dezhi Luo Published: 2026-09-25 arXiv: 2609.28654

Motivation

Object permanence and solidity are hallmarks of human cognitive priors. Recent studies show that video generation models — currently the paradigmatic representative of world models — have begun to exhibit emergent reasoning abilities, making them ideal candidates for building human-like physical intelligence. This work asks: do video models possess emergent object permanence?

WROP Benchmark

WROP (World Reasoning with Object Permanence): a data infrastructure of 150 hand-designed, cognitive-science-inspired tasks
Tasks divided into six cognitive categories
Blender-based generators randomize distractor parameters (speed, lighting, camera angles, etc.) while preserving each task's cognitive structure
Each task yields 10,000+ samples

Releases

A 1.5-million-sample training corpus
A 300-question exam
Data, exam, model answers, scores, model weights, and a native PyTorch training stack — all open-sourced

Evaluation

Fourteen video models were evaluated on the exam, including the authors' own 16B world model, PWM-WROP. In a blind pairwise Elo study, PWM-WROP ranked first among continuation models and third overall.

---

*Auto-collected on 2026-09-26*