With no training, a three-token opening cue takes the Olmo 3 base model to 82 on MATH-500 at AI2's own sampling setting, against 93 for the RL-trained model.