Bridging the Knowledge-Prediction Gap in LLMs on Multiple-Choice Questions