As large language models have already become essential components of so many real-world applications, understanding how they reason and learn from prompts is critical.