ModelRefs / Self-Consistency — AI Glossary

Self-Consistency — AI Glossary

A decoding strategy that samples multiple chain-of-thought reasoning paths and majority-votes the final answer.

Overview

Self-consistency (Wang et al., 2022) yields substantial gains on math and reasoning benchmarks at the cost of N× tokens and latency. Reduces sensitivity to unlucky single-sample reasoning paths.

Reference details

Topic	reasoning
Last reviewed	2026-06-24

How it works

The same prompt is sampled several times at non-zero temperature, so each run takes an independently varying reasoning path.
Only the final answer is extracted from each run; the intermediate reasoning is discarded rather than compared.
The answers are tallied and the most frequent one is returned, so correctness is decided by agreement rather than by any judgement of quality.
No path is evaluated or pruned, which is why the method is simple to implement and why it cannot recover when the model is consistently wrong.
It requires a discrete answer that can be compared for equality, which is what limits it to problems with a definite result.

Example: Why majority voting helps only above chance

If a single sample is right 60% of the time and errors are independent, five samples give a majority right about 68% of the time — the binomial probability of at least three of five successes at p = 0.6. The gain is real and it costs five times the tokens. At p = 0.4 the same vote lands near 32%, below the single sample: majority voting amplifies whatever the model's tendency already is. On a problem where the model is reliably wrong, sampling more makes the wrong answer more certain.

Commonly confused with

Like tree of thoughts, this samples more than one reasoning path — but it never evaluates or prunes them. It generates several independently and takes the majority answer, which makes it far cheaper and means it needs answers that can be compared for equality. That works for a number or a label and does not work for an essay, where there is no majority to take.

When to use it

Reach for it when:

Problems with a single discrete answer that can be compared across samples.
The model is better than chance but single samples are unreliable.
Accuracy justifies several times the tokens and latency.

Reach for something else when:

Open-ended generation, where two correct answers differ in wording and cannot be tallied.
The model is systematically wrong, where voting entrenches the error.
Cost or latency budgets cannot absorb the multiplier.

Primary source

Wang et al., Self-Consistency Improves Chain of Thought Reasoning in Language Models (arXiv)

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Self-Consistency — AI Glossary.

Frequently asked questions

What is Self-Consistency?