MOTIF trained model and Vanilla GRPO trained model, compared in the paper.