The prompt in table 1 makes it very likely that the model will use the correct format. The pretrained model is pretty good so it only needs to stumble upon a co

The prompt in table 1 makes it very likely that the model will use the correct format. The pretrained model is pretty good so it only needs to stumble upon a correct answer every once in a while to start making progress. Some additional details in the Shao et al, 2024 paper.