Self-Consistency
A prompt engineering technique that improves reasoning accuracy by sampling multiple chain-of-thought reasoning paths for the same question and taking the majority-vote result, trading computational cost for reliability.
Core Facts
The core method of Self-Consistency is to independently sample the same problem multiple times, generate multiple different chains of thought, and select the answer with the highest frequency through majority voting.
Self-Consistency was formally proposed by Google Research in the 2022 paper 'Self-Consistency Improves Chain of Thought Reasoning in Language Models.'
The Self-Consistency method generates multiple chains of thought simultaneously and takes the majority answer, but carries the risk of multiple incorrect chains outvoting the correct one.
Timeline (last 90 days)
Test-Time Compute improves answer quality by adding sampling, search, or reflection steps at the inference stage, for example Self-Consistency sampling multiple reasoning paths in parallel and then voting, and Reflexion letting the Agent self-criticize and retry; each sample and each parallel path corresponds to a complete forward computation.
Ensemble sampling methods (such as self-consistency checks) estimate uncertainty by repeatedly sampling the same question and statistically analyzing the answer distribution, achieving higher accuracy but multiplying inference costs.
Prompt engineering refers to the techniques and methodologies for guiding large language models to produce more accurate outputs through carefully designed input prompts, encompassing technical paradigms such as zero-shot prompting, few-shot prompting, chain-of-thought, and self-consistency.
Self-Consistency由Wang等人提出，Tree-of-Thought由Yao等人提出
Self-Consistency由Wang等人提出，通过生成多条推理路径后多数投票选择最一致的答案
高级提示技术包括Tree-of-Thought（思维树）、Self-Consistency（自洽性采样）、ReAct（推理+行动）等，通过设计输入格式在不改变模型参数的情况下激活模型潜在推理能力
All Facts (16)
The core method of Self-Consistency is to independently sample the same problem multiple times, generate multiple different chains of thought, and select the answer with the highest frequency through majority voting.
90%VerifiedSelf-Consistency was formally proposed by Google Research in the 2022 paper 'Self-Consistency Improves Chain of Thought Reasoning in Language Models.'
75%VerifiedThe Self-Consistency method generates multiple chains of thought simultaneously and takes the majority answer, but carries the risk of multiple incorrect chains outvoting the correct one.
65%VerifiedThe essence of the Self-Consistency strategy is to trade computational resources for reasoning reliability.
65%VerifiedSelf-Consistency typically sets the temperature to 0.7 or above to increase sampling diversity.
65%Unverified大模型的规划能力经历了四个进化阶段：基础输入输出、Chain of Thought（思维链）、Self-Consistency（自洽性）、Tree of Thought（思维树）
80%Unverified2023年以来，学术界提出了思维链（Chain-of-Thought）、思维树（Tree-of-Thought）、自我一致性（Self-Consistency）等推理增强技术
75%UnverifiedThe Self-Consistency method, proposed by Xuezhi Wang and others, improved accuracy by more than 10% on mathematical reasoning benchmarks such as GSM8K.
70%UnverifiedOn the GSM8K mathematical reasoning benchmark, Self-Consistency can improve accuracy by approximately 17 percentage points compared to single greedy decoding.
60%UnverifiedSelf-Consistency（自洽性采样）和Tree-of-Thought（思维树）是Chain-of-Thought的后续衍生方法
60%UnverifiedTest-Time Compute improves answer quality by adding sampling, search, or reflection steps at the inference stage, for example Self-Consistency sampling multiple reasoning paths in parallel and then voting, and Reflexion letting the Agent self-criticize and retry; each sample and each parallel path corresponds to a complete forward computation.
50%UnverifiedEnsemble sampling methods (such as self-consistency checks) estimate uncertainty by repeatedly sampling the same question and statistically analyzing the answer distribution, achieving higher accuracy but multiplying inference costs.
50%UnverifiedPrompt engineering refers to the techniques and methodologies for guiding large language models to produce more accurate outputs through carefully designed input prompts, encompassing technical paradigms such as zero-shot prompting, few-shot prompting, chain-of-thought, and self-consistency.
50%UnverifiedSelf-Consistency由Wang等人提出，Tree-of-Thought由Yao等人提出
50%UnverifiedSelf-Consistency由Wang等人提出，通过生成多条推理路径后多数投票选择最一致的答案
50%Unverified高级提示技术包括Tree-of-Thought（思维树）、Self-Consistency（自洽性采样）、ReAct（推理+行动）等，通过设计输入格式在不改变模型参数的情况下激活模型潜在推理能力
50%