Pre-trained language models (LM) have become go-to text representation encoders. Prior research used deep LMs to encode text sequences such as sentences and passages into single dense vector representations. These dense representations have been used in efficient text comparison and em