Self-consistency
Running the same prompt multiple times and taking a vote on the answer, boosting accuracy when reasoning is involved.
Think of it like
Asking the same question to the same person ten times; their answers agree more often for true facts than for guesses.
Example
Ask "What is 7 × 13?" five times with temperature=0.7; four times you get 91, once you get 89. Trust the majority.
How it actually works
Self-consistency amplifies the effect of temperature and sampling. At high temperature, the model samples different reasoning paths; some reach right answers, some wrong ones. Voting on the final answer is surprisingly effective for math and logic. Costs more (5–10× calls), but accuracy can jump 10–30% on hard problems.
For product teams
Improves reliability for critical tasks (medical diagnosis, code review, math) with compute trade-off: 5x cost for much higher confidence.
For engineers
Call model 5–10 times, parse results, count correct answers. Use temperature ≥ 0.5 to induce diversity; temperature=0 gives the same answer every time.
Related
- Sampling — Leverages temperature and sampling diversity.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome