Update README.md

README.md

@@ -14,4 +14,16 @@ P2P is trained on 8,000+ hours of human-annotated gameplay videos. The full data
 
 Our smallest model (150M parameters) can be trained in ~70 hours, and the largest model (1.2B parameters) can be trained in ~140 hours on 8× H100 GPUs.
 
-Please checkout our [website](https://elefant-ai.github.io/open-p2p/) to watch our model play against real human player on Roblox games, and checkout our [github](https://github.com/elefant-ai/open-p2p) for training/inference details.
+Please checkout our [website](https://elefant-ai.github.io/open-p2p/) to watch our model play against real human player on Roblox games, 
+and checkout our [github](https://github.com/elefant-ai/open-p2p) for training/inference details. Our [arxiv paper](https://arxiv.org/abs/2601.04575) is also available. 
+
+If you use our models, please kindly consider citing our paper:
+```bibtex
+@misc{yue2026scaling,
+      title={Scaling Behavior Cloning Improves Causal Reasoning: An Open Model for Real-Time Video Game Playing}, 
+      author={Yuguang Yue and Irakli Salia and Samuel Hunt and Chris Green and Wenzhe Shi and Jonathan J. Hunt},
+      year={2026},
+      eprint={2601.04575},
+      archivePrefix={arXiv},
+      primaryClass={cs.LG}
+}