Google

A specific formatting technique used in few-shot prompts is to employ a special token, such as `####`, to clearly demarcate the final answer of a problem. This helps the Large Language Model learn to structure its output in a consistent and predictable way, separating the final result from the intermediate reasoning steps.

Final Answer Formatting Token

You are given a one-shot prompt designed to teach a model how to solve simple algebra problems. The model is correctly performing the reasoning but is not consistently isolating the final answer, making it difficult to parse the output. Modify the example's answer provided below to include a special formatting token that clearly demarcates the final answer. Use `####` as the special token.

**Original Example:**
Q: If 2x + 3 = 11, what is the value of x?
A: To solve for x, we first subtract 3 from both sides of the equation: 2x = 11 - 3, which simplifies to 2x = 8. Then, we divide both sides by 2 to isolate x: x = 8 / 2. The final answer is 4.

Improving Prompt Structure with a Formatting Token

Based on the provided scenario, what specific, simple modification should the developer make to the examples within the prompt to ensure the final answer is consistently and reliably identifiable for automated extraction? Explain why this modification works.

Improving Automated Answer Extraction

The primary purpose of using a special formatting token, such as `####`, to demarcate the final answer in a few-shot prompt is to enhance the language model's internal step-by-step reasoning capabilities.

An engineer is creating a prompt that includes several examples of a math word problem followed by a step-by-step solution. The goal is for the model to learn this reasoning pattern. However, the model's final answers are often buried within its explanatory text, making them hard to extract automatically. The engineer modifies each example by placing the token `####` immediately before the final numerical answer. Why is this modification an effective strategy?

Learn Before

Related