# What is Self-Consistency?

Also called majority vote sampling.

Self-consistency is a technique that samples several independent reasoning paths for the same question and selects the answer that appears most often, rather than trusting a single generated chain. It exploits the observation that correct answers tend to be reached by many different valid routes, while errors are more scattered across sampled outputs.

The method was introduced as an improvement to chain of thought prompting. Instead of one deterministic pass, the model is sampled multiple times at nonzero temperature, each producing its own reasoning and final answer. The reasoning is discarded and the final answers are aggregated, typically by simple majority vote, sometimes weighted by a confidence or scoring signal.

Its reliability depends on a discrete, comparable answer. Arithmetic results, multiple choice selections, extracted field values, and classification labels aggregate cleanly. Free form paragraphs do not, since two correct answers phrased differently count as disagreement, so applying the technique to open ended text requires a normalization or clustering step first.

Cost scales linearly with the number of samples, and returns diminish. Most reported gains appear within the first several samples, with additional samples adding expense for small improvement. The vote spread also carries information: near unanimous agreement and a narrow plurality are very different signals about how much the answer should be trusted.

A crucial limitation is that consistency is not correctness. A model that is confidently wrong for a systematic reason will be wrong consistently, and the vote will confirm the error. Self-consistency reduces variance in sampling, not bias in the underlying model, so it is not a substitute for verification against external ground truth.

## Key points

- Samples multiple reasoning paths and votes on the answer
- Needs discrete, comparable answers to aggregate
- Cost grows linearly with samples, gains diminish
- Vote spread is a useful confidence signal
- Reduces sampling variance, not systematic error

## In practice

An extraction agent pulls the invoice total from a scanned document five separate times. Four samples return 4,820.00 and one returns 4,320.00, a digit misread. The majority answer is used and the disagreement is logged. On a later document the vote splits three to two, and that low agreement routes the document to human review instead of being written to the ledger.

## Related terms

- [Chain of Thought](/en/glossary/chain-of-thought)
- [Confidence Threshold](/en/glossary/confidence-threshold)
- [Reasoning Model](/en/glossary/reasoning-model)
- [Tree Of Thoughts](/en/glossary/tree-of-thoughts)
- [Deterministic vs Probabilistic Behavior](/en/glossary/deterministic-vs-probabilistic-behavior)

[Back to the AI Glossary](/en/glossary)