🌏 RLVR-World: Training World Models with Reinforcement Learning (NeurIPS 2025)

This is the official code base for the paper RLVR-World: Training World Models with Reinforcement Learning.

Give it a star 🌟 if you find our work useful!

🔥 News

• 🚩 2025.10.28: NeurIPS 2025 camera-ready version is released on arXiv.
• 🚩 2024.09.18: RLVR-World has been accepted by NeurIPS 2025, congrats!
• 🚩 2025.05.26: We release all models and datasets.
• 🚩 2025.05.21: We open-source our training codes.
• 🚩 2025.05.21: Our paper is released on arXiv.

📋 TL;DR

We pioneer training world models through RLVR:

• World models across various modalities (particularly, language and videos) are unified under a sequence modeling formulation;
• Task-specific prediction metrics serve as verifiable rewards directly optimized by RL.

🤗 Models and Datasets

At the moment, we provide the following models and datasets:

Modality	Type	Domain	Name
Language	Dataset	Text game	bytesized32-world-model-cot
Language	World model	Text game	bytesized32-world-model-sft
Language	World model	Text game	bytesized32-world-model-rlvr-binary-reward
Language	World model	Text game	bytesized32-world-model-rlvr-task-specific-reward
Language	Dataset	Web navigation	webarena-world-model-cot
Language	World model	Web navigation	webarena-world-model-sft
Language	World model	Web navigation	webarena-world-model-rlvr
Video	Tokenizer	Robot manipulation	rt1-frame-tokenizer
Video	World model	Robot manipulation	rt1-world-model-single-step-base
Video	World model	Robot manipulation	rt1-world-model-single-step-rlvr
Video	Tokenizer	Robot manipulation	rt1-compressive-tokenizer
Video	World model	Robot manipulation	rt1-world-model-multi-step-base
Video	World model	Robot manipulation	rt1-world-model-multi-step-rlvr

💬 Evaluating Language World Models

See lang_wm:

• Text game state prediction
• Web page state prediction
• Application: Model predictive control for web agents

🎇 Evaluating Video World Models

See vid_wm:

• Robot manipulation trajectory prediction
• Application: Real2sim policy evaluation

🎥 Showcases

🚀 Release Progress

• Video world model with RLVR
• Pre-trained & post-trained video world model weights
• Real2sim policy evaluation with video world models
• Text game SFT data
• Web page SFT data
• Language world model on text games with RLVR
• Language world model on web pages with RLVR
• Post-trained language world model weights
• Web agents with language world models

📜 Citation

If you find this project useful, please cite our paper as:

@inproceedings{wu2025rlvr,
    title={RLVR-World: Training World Models with Reinforcement Learning}, 
    author={Jialong Wu and Shaofeng Yin and Ningya Feng and Mingsheng Long},
    booktitle={Advances in Neural Information Processing Systems},
    year={2025},
}

🤝 Contact

If you have any questions, please contact wujialong0229@gmail.com.

💡 Acknowledgement

We sincerely appreciate the following github repos for their valuable codebase we build upon:

• https://github.com/volcengine/verl
• https://github.com/thuml/iVideoGPT
• https://github.com/kyle8581/WMA-Agents
• https://github.com/cognitiveailab/GPT-simulator
• https://github.com/web-arena-x/webarena
• https://github.com/simpler-env/SimplerEnv