Cross-encoders are neural net models that process a pair of inputs together in a single pass to make a prediction about the relationship between them (like the relevance of one text to another). They are not typically used to produce standalone [[Dense Vectors ("Embeddings")]] or [[Sparse Vectors ("Embeddings")]], but rather to score pairs of inputs directly. These models are exceptional in classification and [[Re-ranker]] tasks for [[Information Retrieval (IR)]].