Hokin/object-permanence-benchmark

Object Permanence Benchmark

The 300-question exam of WROP (World Reasoning with Object Permanence) and 14 video models' answers to it: our own model PWM-WROP, five open-source baselines and eight hosted commercial models. Scores and the human-preference leaderboard live on the project page; this repository holds the questions and the raw generations only.

Abstract

Object permanence is the hallmark of human cognitive priors. Recent studies show that video models, also often referred as world models, have also shown emerged reasoning abilities, making them ideal candidates for building human-like multimodal 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, 4 continuation, 3 reference-to-video, 7 edit, and PWM-WROP, our 16B world model. In a blind pairwise Elo study, PWM-WROP ranks third overall, behind a statistical tie between two commercial reference-to-video systems, and first among continuation models. We release the data, exam, model answers, scores, weights, and PWM, our training stack, a native-PyTorch implementation of the model on AWS Trainium2.

At a glance

Property	Value
Questions	300 (150 tasks × 2 samples, _0000 / _0001)
Models	14 - 1 ours, 5 open-source, 8 commercial
Answers	4,197 mp4 (every model answered all 300; runway-aleph-v2v 297)
Question format	Five files per question, same contract as the training corpus
Reference clips	1280 × 720, 24 fps, 60 input frames → 60 target frames
Size	199 MB (eval/) + 4.4 GB (results/)

Layout

.
├── LICENSE
├── eval/                                  # the exam — 300 questions, 1,500 files
│   └── <shard>/<task>_task/
│       ├── <task>_0000/
│       │   ├── input_video.mp4            # the question: 60 frames the model sees
│       │   ├── target_video.mp4           # the reference answer: 60 frames that follow
│       │   ├── prompt.txt                 # the continuation instruction
│       │   ├── trajectory.npz             # per-frame ground truth for every body
│       │   └── metadata.json
│       └── <task>_0001/                   # same five files
└── results/                               # the answers — 4,197 mp4
    ├── pwm-wrop/<task>_task/<task>_0000.mp4
    ├── ltx-2.3-dev-v2v/…
    └── …                                  # 14 model directories

Usage

huggingface-cli download Hokin/object-permanence-benchmark --repo-type dataset --local-dir ./object-permanence-benchmark

Each answer results/<model>/<task>_task/<sample_id>.mp4 is that model's continuation of eval/<shard>/<task>_task/<sample_id>/input_video.mp4; the reference answer is the matching target_video.mp4.

License

CC BY-NC 4.0 — attribution: Hokin Deng. Non-commercial use only. For commercial licensing contact hokinxqdeng@gmail.com.

Citation

Training Object Permanence in World Models (arXiv link to follow).

@article{zhang2026training,
  title   = {Training Object Permanence in World Models},
  author  = {Zhang, Haotian and others},
  year    = {2026},
  url     = {https://object-permanence.world}
}