“A phenomenon where the model oscillates between correct and incorrect answers due to training rewards.”