How it works
Self-consistency is a decoding strategy that generates multiple answers for the same input, with varied chains of thought, and picks the most frequent answer among them. Instead of a single reasoning path, the model samples several trajectories and follows the majority. Sampling ensures each attempt follows a slightly different path, and a simple rule counts which answers appear most often before returning the most repeated one. The diversity of attempts is controlled by temperature and the number of samples, and the final result is a kind of vote among reasoning trajectories. This process works well on multiple choice questions, step by step reasoning and decisions with several alternatives, as long as each attempt is truly independent. The more valid paths the model finds, the better the majority choice. The majority rule works well when attempts are independent; if they all follow the same flawed reasoning, the result will be too. To mitigate that risk, it is recommended to set a moderate temperature and revisit cases where attempts agree with each other.
Practical example
For a complex sum, the model generates five calculation paths. Four arrive at the same result and one at a slightly different value. The final result is the one that appears most often, which corrects some occasional errors in each path. In multiple choice problems, the answer that appears most among the variations tends to be more reliable, because the model converges to a consensus among several different reasonings.
What it's for
Self-consistency is used to improve response stability in reasoning, math and decision making tasks. It works well when the model has access to several valid trajectories and the error of one of them does not blow up the final result. The cost is more tokens and time, so it is worth using in decisive cases where the price of an error is high. In fast, routine tasks, the benefit may not justify the additional processing cost.
Difference between X and Y
Unlike single chain of thought, which guides the model along a single path, self-consistency uses several. Temperature and the number of samples control the diversity of attempts; a simple chain always follows the most likely path, without varying the reasoning.