Large language models (LLMs) have revolutionized how we interact with machines, but they're not without their quirks. Sometimes, they struggle to follow instructions accurately. Recent research suggests a clever solution: Minimum Bayes Risk (MBR) decoding. Imagine an LLM generating multiple possible responses to a prompt. MBR acts like a discerning judge, evaluating each response against the others. It then selects the response with the highest average 'utility' or quality score, ensuring a higher chance of instruction adherence. This approach leverages the LLM's ability to assess its work, choosing the 'best' answer from a pool of possibilities. Testing this method on popular benchmarks like AlpacaEval and MT-Bench reveals that smaller LLMs can effectively supervise larger ones, improving instruction-following capabilities. Researchers found that LLMs using MBR outperformed those using traditional methods like greedy or beam search decoding, leading to more accurate and reliable outputs. To make MBR more efficient, researchers experimented with 'distilling' the knowledge gained from MBR evaluations back into the model itself. By iteratively refining the model on its own best outputs, as judged by MBR, they aimed to achieve similar gains without the extra computational cost of repeated evaluations. The results were promising: this 'distilled' model performed similarly to the MBR-enhanced LLM while being much faster. This research opens exciting doors for the future of LLMs. By learning to judge and refine their own work, LLMs can become even more effective instruction followers, bridging the gap between AI capabilities and human expectations.