You are interacting with a language model to solve the problem presented in the case study. The model has provided an initial, incorrect answer. Your task is to write the next single prompt you would use to guide the model to evaluate its own work and find the correct solution, without giving it the answer directly.

Google

Following the initial problem-solving attempt in a multi-round interaction, the second step involves prompting the Large Language Model to assess the accuracy of its own answer. If the initial solution is found to be incorrect, the model is then instructed to revise its work to produce an improved solution.

Second Step in Multi-Round Interaction: Answer Evaluation and Refinement

During the second stage of a multi-round interaction, a prompt template is used to instruct a Large Language Model to assess and improve a previously generated solution. This template presents both the original question and the initial response to the model. An example prompt is: "You will be provided with a math problem, along with a solution. Evaluate the correctness of this solution, and identify any errors if present. Then, work out your own solution. Problem: $${*\mathrm{problem}*}$$ Solution: $${*\mathrm{solution}*}$$".

Prompt Template for Answer Evaluation and Refinement

A user asks a language model: 'What is the total number of days in the first six months of a non-leap year (January to June)?' The model incorrectly responds: 'There are 180 days.' The user wants to guide the model to the correct answer. Consider two possible follow-up prompts:

**Approach A:** "That doesn't seem right. Please try calculating the total number of days in the first six months of a non-leap year again."

**Approach B:** "Please review your previous answer. First, list the number of days in each of the first six months. Then, sum those numbers to provide a final, corrected total."

Evaluate these two approaches. Which one is more likely to lead to a correct and reliable answer from the model? Justify your choice by explaining the difference in the reasoning process each prompt encourages.

Evaluating Model Correction Strategies

Guiding a Model to Self-Correct

You are designing a conversational interaction to help a language model solve a complex problem more reliably. Arrange the following steps into the correct logical sequence for a two-round interaction focused on self-correction.

Learn Before

Related