add paper link

README.md

@@ -9,7 +9,7 @@ tags:
 license: mit
 ---
 DCFormer-2.8B is a pretrained language model on the Pile with 300B tokens, which is a parameter and computation efficient attention architecture that tackles the shortcomings of MHA
-and increases the expressive power of the model by dynamically composing attention heads. It is short for DCFormer++2.8B and please see downstrem evaluations and more details in the paper[(Improving Transformers with Dynamically Composable Multi-Head Attention)](). In addition, we open-source Jax training code on [(Github)](https://github.com/Caiyun-AI/DCFormer/).
+and increases the expressive power of the model by dynamically composing attention heads. It is short for DCFormer++2.8B and please see downstrem evaluations and more details in the paper[(Improving Transformers with Dynamically Composable Multi-Head Attention)](https://arxiv.org/abs/2405.08553). In addition, we open-source Jax training code on [(Github)](https://github.com/Caiyun-AI/DCFormer/).
 
 We recommend <strong>compiled version</strong> of DCFormer with *torch.compile* for inference acceleration. Please refer to Generation section for compile implementation.