Datasets:
metadata
license: apache-2.0
task_categories:
- video-classification
language:
- en
tags:
- reward-model
- world-model
- video-generation
- camera-control
- preference-data
size_categories:
- n<1K
configs:
- config_name: default
data_files: bench.jsonl
WorldReward-Bench
A human-annotated preference benchmark for camera-conditioned world models. 760 pairs of videos, each pair generated by two different models from the same source image and the same camera-action sequence, with human verdicts on three independent axes.
- 📰 Paper: https://arxiv.org/abs/2609.03952
- 🪐 Project Page: https://codegoat24.github.io/WorldReward
- 🤗 Model Collections: https://huggingface.co/CodeGoat24/WorldReward-9B
- 🚀 Github: https://github.com/CodeGoat24/WorldReward
- 👋 Point of Contact: Yibin Wang
Evaluation Dimensions
| Axis | Question |
|---|---|
action |
Did the camera actually execute the commanded motion — right direction, right magnitude, no un-commanded drift? |
appearance |
Which video looks better — fewer artifacts, more stable structure, faithful to the source scene? |
motion |
Which video is genuinely generating new content, rather than sliding a static texture or melting? |
Each verdict is left, right, or tie.
Contents
bench.jsonl # 760 pairs, one JSON object per line
videos/<pair_id>/source.* # shared source image (exact path is in bench.jsonl)
videos/<pair_id>/left.mp4 # the video shown on the left
videos/<pair_id>/right.mp4 # the video shown on the right
videos/<pair_id>/left_overlay.mp4 # left.mp4 with the commanded action burned in
videos/<pair_id>/right_overlay.mp4 # right.mp4 with the commanded action burned in
The *_overlay.mp4 files are a visualisation aid for inspecting trajectories by
eye.
Source images keep their original format and pixels when they are at most 2048px on the long side. Larger ones are downscaled to 2048px and saved as JPEG.
Schema
| Field | Type | Description |
|---|---|---|
pair_id |
string | Opaque identifier, wrb_0001-style. Carries no metadata. |
input_image |
path | Source image both videos were generated from |
input_caption |
string | English description of the source scene |
actions |
list[string] | Commanded camera-action sequence, one token per step |
frames_per_action |
int | Frames each action occupies |
num_frames |
int | Total frames per video |
left / right |
object | video, overlay, and the generating model name |
trajectory_family |
string | One of 9 fine-grained trajectory types |
trajectory_group |
string | pure_translation / pure_rotation / compound |
style |
string | photo / game_anime / traditional |
label |
object | Human verdicts: action, appearance, motion, each left/right/tie |
Results
Three-way agreement with the human labels (%). All 760 pairs count; a pair whose
label is tie is correct only if the model also predicts tie. Act./App./Mot. =
action / appearance / motion. Best and second-best per column; -- marks an
axis a predictor does not model.
| Reward model | All Act. |
All App. |
All Mot. |
Translation Act. |
Translation App. |
Translation Mot. |
Rotation Act. |
Rotation App. |
Rotation Mot. |
Compound Act. |
Compound App. |
Compound Mot. |
Photo Act. |
Photo App. |
Photo Mot. |
Game/Anime Act. |
Game/Anime App. |
Game/Anime Mot. |
Art Act. |
Art App. |
Art Mot. |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Closed-source VLM | |||||||||||||||||||||
| Gemini-3.1-Pro | 65.79 | 80.13 | 60.79 | 64.73 | 82.19 | 64.38 | 62.90 | 83.87 | 57.53 | 68.79 | 75.53 | 59.22 | 65.48 | 82.74 | 63.84 | 65.27 | 79.34 | 57.49 | 70.49 | 68.85 | 60.66 |
| GPT-5.5 | 74.21 | 79.87 | 69.47 | 72.95 | 81.16 | 74.66 | 72.04 | 84.41 | 62.90 | 76.95 | 75.53 | 68.44 | 76.44 | 81.64 | 68.77 | 71.56 | 79.64 | 68.86 | 75.41 | 70.49 | 77.05 |
| Image / video quality reward models | |||||||||||||||||||||
| VideoAlign | -- | 61.32 | 40.13 | -- | 66.44 | 31.51 | -- | 60.22 | 45.16 | -- | 56.74 | 45.74 | -- | 61.10 | 41.10 | -- | 61.08 | 41.32 | -- | 63.93 | 27.87 |
| UnifiedReward-Flex | -- | 64.62 | 49.86 | -- | 63.90 | 46.57 | -- | 73.18 | 56.98 | -- | 59.78 | 48.55 | -- | 65.08 | 55.31 | -- | 63.64 | 47.02 | -- | 67.27 | 30.91 |
| UnifiedReward-Think | -- | 66.09 | 38.79 | -- | 65.41 | 32.88 | -- | 69.73 | 52.43 | -- | 64.41 | 35.94 | -- | 69.51 | 41.21 | -- | 64.37 | 37.43 | -- | 55.00 | 31.67 |
| Aesthetic | -- | 69.87 | -- | -- | 66.10 | -- | -- | 72.04 | -- | -- | 72.34 | -- | -- | 66.58 | -- | -- | 75.15 | -- | -- | 60.66 | -- |
| HPSv3 | -- | 73.68 | -- | -- | 74.66 | -- | -- | 76.34 | -- | -- | 70.92 | -- | -- | 73.15 | -- | -- | 75.45 | -- | -- | 67.21 | -- |
| Geometry estimation models | |||||||||||||||||||||
| DAv3 | 70.53 | -- | -- | 67.47 | -- | -- | 68.82 | -- | -- | 74.82 | -- | -- | 70.96 | -- | -- | 75.75 | -- | -- | 39.34 | -- | -- |
| WorldMirror | 68.55 | -- | -- | 67.81 | -- | -- | 68.28 | -- | -- | 69.50 | -- | -- | 67.40 | -- | -- | 74.25 | -- | -- | 44.26 | -- | -- |
| Backbone, zero-shot | |||||||||||||||||||||
| Qwen3.5-VL-9B | 48.42 | 48.29 | 43.82 | 48.29 | 45.55 | 41.78 | 52.69 | 52.15 | 49.46 | 45.74 | 48.58 | 42.20 | 50.14 | 43.84 | 47.12 | 47.01 | 52.10 | 42.81 | 45.90 | 54.10 | 29.51 |
| Qwen3.5-VL-27B | 63.68 | 44.34 | 62.76 | 65.07 | 37.33 | 65.75 | 65.05 | 51.61 | 63.44 | 61.35 | 46.81 | 59.22 | 64.93 | 38.36 | 66.85 | 62.87 | 51.50 | 58.68 | 60.66 | 40.98 | 60.66 |
| WorldReward-9B | 77.63 | 81.32 | 73.03 | 76.71 | 77.74 | 78.77 | 73.12 | 86.02 | 64.52 | 81.56 | 81.91 | 72.70 | 77.26 | 81.37 | 71.78 | 78.74 | 82.04 | 75.75 | 73.77 | 77.05 | 65.57 |
Usage
See https://github.com/CodeGoat24/WorldReward for the full evaluation protocol.
Citation
@article{WorldReward,
title={WorldReward: Reward Modeling for Camera-Conditioned World Models},
author={Wang, Yibin and Wang, Zehan and Tang, Junshu and Li, Zhimin and Zhou, Yujie and Bu, Jiazi and Ling, Pengyang and Han, Feng and Zhang, Zhixiong and Xing, Long and Ding, Shengyuan and Li, Ziang and Jin, Cheng and Zang, Yuhang and Wang, Jiaqi and Pang, Tianyu},
journal={arXiv preprint arXiv:2609.03952},
year={2026}
}