在本節中，我們將超越基本提示，深入理解推理的數學引擎。我們探討如何 鏈式思維（CoT） 透過利用 Transformer 的上下文視窗作為外化的工作記憶，將大型語言模型從單純的查表工具轉化為順序推理引擎。
1. 從直接映射到順序機率
在標準提示中，模型僅根據問題計算答案的機率——$P(答案 | 問題)$。對於複雜邏輯，這種單一「跳躍」往往過大。鏈式思維則將其分解為一系列步驟：$P(y | x, z_1, z_2, \dots, z_n)$，其中每個 $z$ 均為 中間推理步驟 ，這些步驟能約束模型朝向更準確的結論前進。
2. 邏輯透明性與運算
CoT 本質上是一種「放慢」模型速度的方法。透過強制生成中間令牌，我們迫使模型將更多運算週期投入於問題之中。這產生了一條「可驗證的痕跡」，使模型的內部處理過程轉化為可審計、可修正且可結構化的清晰路徑。
手動藍圖結構
[背景/任務]限制條件： 逐步推導的要求
推理藍圖：
1. 步驟 1
2. 步驟 2
3. 步驟 3
[最終輸出格式]
Type a message... (Disabled in Demo Mode)
Question 1
What is the primary mechanical function of intermediate tokens ($z$-tokens) in a Chain of Thought?
Challenge: The Logical Intervention
Fixing a failing model.
Scenario: A model is consistently failing a multi-step math word problem by providing a wrong answer immediately. Your goal is to move beyond the "Think step-by-step" trigger.
Strategy
How can you use a "manual blueprint" to ensure the model doesn't skip critical logic in its context window?
Solution:
You must define a specific structure in the prompt, such as "Identify Constants -> Define Variables -> Perform Operations -> State Result," which forces the model to externalize each dependency before the final prediction.
You must define a specific structure in the prompt, such as "Identify Constants -> Define Variables -> Perform Operations -> State Result," which forces the model to externalize each dependency before the final prediction.