Spiral-Qwen3-4B

Links

• 📜 Paper
• 💻 GitHub
• 🤗 Spiral Collection

Introduction

This model is trained with self-play on multi-games (TicTacToe, Kuhn Poker, Simple Negotiation) using the SPIRAL framework.

Citation

@article{liu2025spiral,
  title={SPIRAL: Self-Play on Zero-Sum Games Incentivizes Reasoning via Multi-Agent Multi-Turn Reinforcement Learning},
  author={Liu, Bo and Guertler, Leon and Yu, Simon and Liu, Zichen and Qi, Penghui and Balcells, Daniel and Liu, Mickel and Tan, Cheston and Shi, Weiyan and Lin, Min and Lee, Wee Sun and Jaques, Natasha},
  year={2025},
  journal={arXiv preprint arXiv:2506.24119},
  url={https://arxiv.org/abs/2506.24119}
}