2/ All objectives derived within the LDM framework consist of (i) an alignment term that maximizes log-likelihood under the assumed latent model and (ii) a uniformity term that maximizes entropy of the embeddings to prevent representational collapse.