A preprint finds that reweighting tokens surprising to an AI student can shift math results, but gains vary by model size and task.