Learned positional embeddings are a technique used in transformer models to provide information about the position of tokens in a sequence, allowing the model to capture the order of words. Unlike fixed positional encodings, learned embeddings are trainable parameters that can adapt to the specific data and task, potentially improving model performance.