Latent Positional Information is in the Self-Attention Variance of Transformer Language Models Without Positional Embeddings ACL

Ta-Chung Chi | Ting-Han Fan | Li-Wei Chen | Alexander Rudnicky | Peter Ramadge |