Training Object Permanence in World Models

arXiv 2026 33.4 method

TLDR

Introduces WROP, a 1.5M-sample synthetic dataset and exam to train and evaluate object permanence in video generation world models; PWM-WROP ranks top among continuation models.

Reasoning

The paper presents a large, carefully designed synthetic benchmark and evaluates many video models, with strong resource release and empirical rigor. However, the reliance on Blender-generated synthetic tasks may limit generalization to real-world physical scenes, and the abstract does not demonstrate interactive or RL-based world model capabilities.

Read-first score

Read-first score 33.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 36.

Recency 5%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 15%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=dataset,evaluation

Topical relevance 25%

51.4

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 15%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=dataset

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 10%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 258.

Keyword Scores

world model

video world model

generative world model

world dynamics prediction

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

WROP data infrastructure with 150 hand-designed cognitive science inspired tasks divided into six cognitive categories for object permanence and solidity.
Blender-based generators that randomize speed, lighting, camera angle, and other nuisance parameters while preserving each task's cognitive structure, yielding 10,000+ samples per task.
Release of a 1.5M-sample training corpus and a 300-question exam for evaluating object permanence in video models.
PWM-WROP, a 16B world model, and PWM, a native-PyTorch training stack on AWS Trainium2.
Blind pairwise Elo evaluation across 14 video models spanning reference-to-video, edit, and continuation tasks.

Methodology

The authors designed 150 cognitive-science-inspired tasks across six categories and built Blender generators that randomize nuisance parameters such as speed, lighting, and camera angle while preserving cognitive structure. This produced a 1.5M-sample training corpus and a 300-question exam. They evaluated 14 video models, including their 16B PWM-WROP, using a blind pairwise Elo study.

Key Results

On the 300-question WROP exam, PWM-WROP ranks first among continuation models and third overall, behind only a statistical tie between two reference-to-video models.

Resources

Paper