## Diagram: Comparison of Chain-of-Thought (CoT) Prompt Structures
### Overview
The image is a technical diagram illustrating three different structures for Chain-of-Thought (CoT) prompting, used to guide AI models in reasoning tasks. It visually compares "Few-shot CoT prompt w/o textual instructions," "Few-shot CoT prompt w/ textual instructions," and "Zero-shot CoT prompt." The diagram uses color-coded boxes and labels to break down the components of each prompt type.
### Components/Axes
The diagram is organized into three main rectangular boxes arranged horizontally, each representing a prompt type.
**1. Left Box: "Few-shot CoT prompt w/o textual instructions"**
* **Title:** "A CoT prompt for Mathematical Reasoning"
* **Structure:** Divided into three labeled sections:
* **Question** (Blue label, top-left): Contains the user's query.
* **Rationale** (Orange label, middle-left): Contains the step-by-step reasoning.
* **Answer** (Green label, bottom-left): Contains the final answer.
* **Content:**
* **Question:** "Q: Leah had 32 chocolates and her sister had 42. If they ate 35, how many pieces do they have left in total?"
* **Rationale:** "A: Originally, Leah had 32 chocolates and her sister had 42. So in total they had 32 + 42 = 74. After eating 35, they had 74 - 35 = 39 pieces left in total."
* **Answer:** "The answer is 39."
* **Query Section:** Below the main box, a separate purple-bordered box labeled "Query" contains a new problem: "Q: If there are 3 cars in the parking lot and 2 more cars arrive, how many cars are in the parking lot?"
**2. Middle Box: "Few-shot CoT prompt w/ textual instructions"**
* **Title:** "A CoT prompt for Date Reasoning"
* **Structure:** Same three-section layout (Question, Rationale, Answer).
* **Key Difference:** A green dashed arrow labeled **"Textual Instructions"** points from the right side of this box to the rationale text, highlighting the inclusion of explicit reasoning instructions.
* **Content:**
* **Question:** "Q: It is 4/19/1969 today. What is the date 24 hours later in MM/DD/YYYY?"
* **Rationale:** "A: Today is 04/19/1969. **24 hours later** is one day after today, which would be 04/20/1969. **Let's think step by step.**" (The bolded phrases are the "textual instructions").
* **Answer:** "The answer is 04/20/1969."
* **Query Section:** Below, a purple-bordered "Query" box contains: "Q: Today is May 25 2013. What is the tomorrow of today in MM/DD/YYYY? **Let's think step by step.**"
**3. Right Box: "Zero-shot CoT prompt"**
* **Title:** "A CoT prompt for Commonsense Reasoning"
* **Structure:** A single, simpler box containing only a question and the instruction "Let's think step by step." It lacks the separate Rationale and Answer sections shown in the few-shot examples.
* **Content:**
* **Question:** "Q: Who lived longer, Theodor Haeker or Harry Vaughan Watkins?"
* **Instruction:** "Let's think step by step."
### Detailed Analysis
* **Spatial Grounding:** The "Demonstrations" label (blue) is positioned to the far left, indicating the first two boxes serve as examples. The "Query" boxes are consistently placed below their respective demonstration boxes. The "Textual Instructions" label and arrow are in the upper-right quadrant, specifically referencing the middle box.
* **Trend Verification:** The diagram demonstrates a progression in prompt structure:
1. **Few-shot w/o instructions:** Provides complete example demonstrations (Q, Rationale, A) but no explicit directive on *how* to reason.
2. **Few-shot w/ instructions:** Provides the same complete demonstrations but embeds the key phrase "Let's think step by step" within the rationale, explicitly modeling the desired reasoning process.
3. **Zero-shot:** Provides no worked examples. It only gives the target question followed by the imperative instruction "Let's think step by step," relying on the model's inherent capability to generate a reasoning chain.
* **Component Isolation:**
* **Header Region:** Contains the titles for each prompt type.
* **Main Chart Region:** Contains the three core prompt structure boxes and their internal Q/R/A sections.
* **Footer Region:** Contains the "Query" boxes for the first two types, showing how the learned structure is applied to a new problem.
### Key Observations
1. **Structural Evolution:** The diagram clearly shows the architectural difference between few-shot learning (which provides examples) and zero-shot learning (which does not).
2. **Role of Textual Instructions:** The middle box highlights that adding the simple phrase "Let's think step by step" to the demonstration rationale is a key method for "textual instructions," explicitly teaching the model the desired behavior.
3. **Domain Variation:** The examples cover different reasoning domains: arithmetic (mathematical), calendar calculation (temporal), and biographical comparison (commonsense), suggesting the CoT method's broad applicability.
4. **Visual Coding:** Color is used functionally: blue for questions, orange for rationales, green for answers, and purple for new queries. Dashed lines group related components.
### Interpretation
This diagram serves as a technical schematic for designing effective prompts to elicit reasoning from AI models. It illustrates a core principle in prompt engineering: **explicitly demonstrating the desired intermediate reasoning steps ("Rationale") significantly structures the model's output.**
The progression from left to right shows increasing reliance on the model's internal capabilities versus external guidance. The "Few-shot w/o textual instructions" prompt relies on pattern matching from examples. The "Few-shot w/ textual instructions" prompt adds a crucial layer of explicit meta-instruction, making the desired process unambiguous. The "Zero-shot" prompt removes the safety net of examples entirely, testing if the instruction alone is sufficient to trigger the reasoning chain.
The inclusion of different domains (math, dates, commonsense) implies that the CoT prompting framework is a general-purpose technique for improving performance on tasks requiring multi-step logic, not just a solution for a single problem type. The "Query" sections demonstrate the practical application: after seeing the structured demonstrations, the model is expected to apply the same Q->Rationale->Answer format to novel problems.