Those are multiple-choice questions. If some of them are "trick questions", where obvious answer (e.g. the value taken directly from question's text) is wrong, bad model might perform worse than a dice.

On the other hand, not sure where from 25% baseline for random answers come from. Since this is multiple-choice-out-of-4 test, random guessing should be correct in 1 in 15 cases, not 1 in 4.