Definition
Loss aggregation is how individual per-example losses get combined into the scalar gradient signal a training step actually uses — a sum, a mean, a token-weighted average, a length-normalized score. The choice quietly biases what behaviors get reinforced, especially in RL and preference training.