Objective minimized during training, e.g., cross‑entropy for classification or language modeling.