Large language models (LLMs) have revolutionized how we approach complex tasks, but their massive size often presents deployment challenges. Distillation, the process of transferring knowledge from a larger "teacher" model to a smaller "student" model, offers a solution. However, the effectiveness of distillation depends on how the knowledge is transferred. Injecting "chains of thought" (CoT), essentially reasoning steps, can significantly boost this process—but how, exactly, does it work? New research dives deep into this intriguing area. Researchers explored CoT-augmented distillation by experimenting with different ways CoT sequences were used in training smaller language models. One surprising finding: the order of information matters! Placing the CoT *after* the target answer during training consistently outperformed traditional placement *before* the answer. What's even more intriguing, shuffling the CoT sequence had little impact on performance when placed after the answer. This suggests that it's not the logical flow of reasoning in the CoT, but the presence of key contextual tokens that enhances the learning process. Digging further, researchers pinpointed crucial tokens within CoTs using a technique called gradient attribution. By including just these essential tokens during distillation, they achieved similar performance improvements as using the full CoT. What does this all mean? It hints at a potential shortcut in knowledge transfer. Instead of mimicking the entire reasoning process, student models might just need to grasp the core concepts linking the question to the correct answer. This opens exciting avenues for optimizing LLM distillation, potentially leading to smaller, more efficient models that retain the reasoning prowess of their larger counterparts.