WorldReward-qwen35-9b

WorldReward: Reward Modeling for Camera-Conditioned World Models.

• 📰 Paper: https://arxiv.org/abs/2609.03952
• 🪐 Project Page: https://codegoat24.github.io/WorldReward
• 🤗 Model Collections: https://huggingface.co/CodeGoat24/WorldReward-qwen35-9b
• 🚀 Github: https://github.com/CodeGoat24/WorldReward
• 👋 Point of Contact: Yibin Wang

Usage

git clone https://github.com/CodeGoat24/WorldReward
cd WorldReward && pip install -e .

python examples/run_single_pair.py \
    --input-image  my_data/scene.jpg \
    --left-video   my_data/system_x.mp4 \
    --right-video  my_data/system_y.mp4 \
    --caption      "A sunlit street lined with colorful European-style buildings." \
    --actions      forward,forward,left+camera_down \
    --frames-per-action 8 \
    --show-reasoning

Inference needs vLLM new enough to register Qwen3_5ForConditionalGeneration:

python -c "from vllm.model_executor.models.registry import ModelRegistry as R; \
           print('Qwen3_5ForConditionalGeneration' in R.get_supported_archs())"

Results

Three-way agreement with human labels on WorldReward-Bench (760 pairs, %). All pairs count: a pair labelled tie is correct only if the model also predicts tie.

Citation

@article{wang2026worldreward,
  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 others},
  journal={arXiv preprint arXiv:2609.03952},
  year={2026}
}