README.md

metadata

license: apache-2.0
datasets:
  - yxzhang2024/SC2-Dynamics-50K
language:
  - en
base_model:
  - Qwen/Qwen3-8B
tags:
  - StarCraft II
  - Dynamics
  - World_Model
  - SC2

StarWM: An Action-Conditioned World Model for StarCraft II

StarWM is the first action-conditioned world model for StarCraft II. Given a structured observation and a sequence of actions, StarWM predicts future observations under partial observability.

📄 Paper:
World Models for Policy Refinement in StarCraft II (arXiv:2602.14857)

🔗 GitHub:
https://github.com/yxzzhang/StarWM

📂 Training Dataset:
SC2-Dynamics-50K

📖 Model Description

StarWM is trained via supervised fine-tuning on Qwen3-8B using the SC2-Dynamics-50K dataset. The model learns to predict textual observations 5 seconds into the future, conditioned on:

• Current textual observation
• A sequence of actions

The textual representation factorizes StarCraft II observation into five semantic modules:

1. Info: Describes economy and status (Minerals, Gas, Collection Rate, Supply, Alerts, Upgrades)
2. Queue: Records ongoing tasks (construction, production, upgrades) and their progress
3. My Units: Includes self units’ IDs, positions, health percentage (HP), energy and status
4. My Structures: Describes self static assets
5. Visible Hostiles: Includes visible enemy units, structures, and snapshot enemy structures under the fog of war

This design enables dynamics modeling in a hybrid and partially observable RTS environment.

🎯 Purpose

StarWM is designed for:

• Action-conditioned observation prediction in StarCraft II
• Short-horizon predictive simulation
• World-model-augmented policy refinement
• Inspiring future research on LLM-based dynamics modeling in complex RTS environments and model-based reasoning under partial observability

📚 Citation

If you use this model, please cite:

@misc{zhang2026worldmodels,
      title={World Models for Policy Refinement in StarCraft II}, 
      author={Yixin Zhang and Ziyi Wang and Yiming Rong and Haoxi Wang and Jinling Jiang and Shuang Xu and Haoran Wu and Shiyu Zhou and Bo Xu},
      year={2026},
      eprint={2602.14857},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2602.14857}, 
}