← AI Terminology
Self-Consistency
Self-consistency samples multiple independent reasoning paths for the same question and selects the answer that appears most often (majority vote over final answers).
Introduced as a decoding strategy that boosts chain-of-thought reliability without extra training.
Introduced as a decoding strategy that boosts chain-of-thought reliability without extra training.
Why It Matters in AI
A single chain-of-thought can be brittle; sampling many and voting cancels uncorrelated errors. It is a simple test-time compute method that often lifts math and logic accuracy significantly.
Key Points
| Aspect | Description |
|---|---|
| Cost | k× inference; parallelisable |
| Limits | Shared systematic biases are not fixed by voting |
| Method | Sample k CoTs at temperature > 0; majority vote on answers |
| Origin | Wang et al., Self-Consistency Improves CoT (2022) |
| Related | Best-of-N with a reward model instead of majority |
| Works best | Tasks with a discrete final answer to vote on |
Simple Analogy
Asking several colleagues to solve a problem independently and going with the answer the most of them got — wisdom of the crowd for model samples.
Common Usage Examples
- k=5–40 CoT samples then mode of answers
- GSM8K/MATH gains over greedy CoT
- Ensemble temperature sweeps
- Combine with tools: vote on code execution results
Summary
In short: Self-consistency samples many reasoning traces and majority-votes the final answer — a simple, strong use of extra inference compute.