Prompting & In-Context Learning

self-consistency

Self-consistency is a way to make chain-of-thought reasoning more reliable by not trusting a single attempt. Instead of asking once, you ask the model to solve the same problem several times, letting it take a slightly different reasoning route each time, and then you keep the final answer that comes up most often. It is the wisdom-of-the-crowd idea applied to one model: many independent tries, then a vote.

The intuition is that a hard problem has many valid paths to the right answer but lots of scattered, idiosyncratic ways to be wrong. Correct paths tend to converge on the same conclusion, while errors disagree with each other, so the most common answer is usually the sound one. To get varied paths you sample with some randomness rather than always taking the single likeliest continuation.

The cost is obvious: solving a problem ten times is roughly ten times the compute and money of solving it once. So self-consistency is worth it for high-stakes questions where accuracy beats speed, and overkill for casual ones. It only helps when there is a clean final answer to tally, such as a number or a choice, and less so for open-ended writing.

Also called
self-consistency decodingmajority vote over reasoning