提示層級架構：從指令到邏輯
提示技術已從簡單的指令式輸入，演進為複雜的推理架構，用以引導模型內部的處理路徑。
核心概念
- 零樣本提示（Zero-shot Prompting）： 在不提供任何範例的情況下，僅提供任務描述（例如：「將這段翻譯成法語」）。
- 少樣本提示（Few-shot Prompting）： 利用「示範」（輸入-輸出對）來定義標籤空間與期望的格式。
- 鏈式思考（Chain-of-Thought, CoT）： 一種促使模型產出中間推理步驟的提示技巧。
- 突現特性： 複雜的推理並非明確編程，而是在參數量通常超過 100 億的模型中「突現」出來。
推理模式的轉變
- 指令遵循：直接將輸入映射至輸出。
- 上下文學習： 從提供的範例中學習模式（少樣本提示）。
- 邏輯分解： 將問題拆解為連續的步驟（鏈式思考）。
- 過程監督： 重視「思考」步驟的準確性，而非僅關注最終答案（如 OpenAI o1 所見）。
關鍵洞察
在少樣本情境中，模型表現高度依賴於標籤分佈與示範的相關性，而不僅僅取決於範例數量。
TERMINALbash — 80x24
> Ready. Click "Run" to execute.
>
Question 1
Which method relies on providing "demonstrations" to guide the model?
Question 2
True or False: Chain-of-Thought reasoning is a capability found in almost all AI models regardless of size.
Challenge: Optimizing Logic Puzzles
Scenario: Optimize a prompt for a model that is struggling with a logic puzzle.
You are using an LLM to solve the following puzzle: "A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost?"
Currently, you are passing the prompt exactly as written above, and the model incorrectly answers "$0.10".
Currently, you are passing the prompt exactly as written above, and the model incorrectly answers "$0.10".
Task 1
Identify if the current prompt is Zero-shot or Few-shot.
Solution:
The current prompt is Zero-shot because it provides the task description without any prior examples or demonstrations of similar solved puzzles.
The current prompt is Zero-shot because it provides the task description without any prior examples or demonstrations of similar solved puzzles.
Task 2
Inject the Zero-shot CoT trigger phrase to improve reasoning accuracy. Rewrite the prompt.
Solution:
"A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost? Let's think step by step:"
"A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost? Let's think step by step:"