A starting value that controls randomness during model training; using the same seed allows researchers to reproduce identical training runs and compare how different seeds affect final model behavior.