Self-Consistency
An LLM reasoning enhancement method proposed by Wang et al. in 2022 that generates multiple independent reasoning chains for the same question and derives a final answer through majority voting or probability aggregation, substantially improving accuracy on arithmetic and commonsense reasoning tasks.
Core Facts
Timeline (last 90 days)
Self-consistency refers to sampling multiple reasoning paths for the same question and taking the answer that appears most frequently; it requires no training and only consumes inference compute, and is often used as a low-cost baseline before reinforcement learning.
Cheap metrics that can be used to detect model drift include self-consistency, output length distribution, character n-gram overlap (Jaccard similarity), error rate, and refusal rate.
Self-consistency typically samples 3-5 reasoning paths and takes a majority vote; the cost is several times that of a single call, and it is suitable for math and logic problems with a unique correct answer.
In practice it is recommended to first use meta-prompting to clarify the requirements, then use self-consistency to handle key calculations, and enable Tree of Thoughts only when exploratory decisions are truly needed, in order to avoid unnecessary token consumption and latency.
All Facts (5)
The methodological system of prompt engineering includes techniques such as zero-shot prompting, few-shot prompting, Chain-of-Thought prompting, and Tree of Thoughts.
70%UnverifiedSelf-consistency refers to sampling multiple reasoning paths for the same question and taking the answer that appears most frequently; it requires no training and only consumes inference compute, and is often used as a low-cost baseline before reinforcement learning.
50%UnverifiedCheap metrics that can be used to detect model drift include self-consistency, output length distribution, character n-gram overlap (Jaccard similarity), error rate, and refusal rate.
50%UnverifiedSelf-consistency typically samples 3-5 reasoning paths and takes a majority vote; the cost is several times that of a single call, and it is suitable for math and logic problems with a unique correct answer.
50%UnverifiedIn practice it is recommended to first use meta-prompting to clarify the requirements, then use self-consistency to handle key calculations, and enable Tree of Thoughts only when exploratory decisions are truly needed, in order to avoid unnecessary token consumption and latency.
50%