Which prompting technique encourages exploration before choosing a most suitable response?

Self-consistency is a technique that involves prompting a model multiple times to generate a diverse set of reasoning paths or outputs for the same question. It then selects the most consistent answer from this set, often through a majority vote. This process inherently encourages the exploration of different possible solutions before the most suitable response is chosen. This contrasts with other methods that typically generate a single, direct answer.

Zero-shot prompting provides no examples and relies on the model's single, initial response without any structured exploration.

Instruction prompting is the general method of giving a direct command, not a specific technique for exploring multiple responses.

Role prompting assigns a persona to the model but does not inherently involve a mechanism for exploring and selecting from multiple final answers.

1. Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., & Zhou, D. (2023). Self-Consistency Improves Chain of Thought Reasoning in Language Models. Section 2, "Self-Consistency," describes the method as sampling "a diverse set of reasoning paths" and then aggregating the results to find the most consistent answer. (https://doi.org/10.48550/arXiv.2203.11171)

2. Wei, J., et al. (2022). Chain-of-thought prompting elicits reasoning in large language models. This foundational paper, which self-consistency builds upon, focuses on generating a single reasoning path, highlighting how self-consistency adds the exploration component. (https://doi.org/10.48550/arXiv.2201.11903)