Self-Consistency Prompting

Self-consistency prompting generates multiple independent reasoning chains for the same question and selects the most frequently occurring answer. It treats the consensus across several reasoning attempts as more reliable than any single chain. Introduced in a 2022 Google Brain paper, it improves accuracy on arithmetic, commonsense reasoning, and logical tasks by 10-30% compared to single chain-of-thought prompting.

The manual approach: ask the same question 3-5 times, each time asking the AI to show its working. Compare the final answers. If the majority agree, that is your answer. If they diverge, the question is ambiguous or the task is at the edge of the model's reliable knowledge — worth verifying externally. In automated systems, you run multiple completions with higher temperature settings and implement majority voting in code.

Self-consistency provides the largest gains on tasks with a definitive correct answer that the model can reach via multiple reasoning paths: arithmetic problems, multi-step logical puzzles, commonsense reasoning, and symbolic manipulation. It helps less for open-ended tasks (creative writing, strategy) where there is no single correct answer. The improvement is largest when the model's single-run accuracy is in the 50-80% range — the regime where it sometimes gets it right but not reliably.

Self-consistency generates multiple independent reasoning chains and takes the majority vote answer — best for tasks with a correct answer. Tree of Thoughts explicitly constructs a search tree, evaluating and pruning branches at each step — best for tasks requiring strategic exploration of a solution space. Self-consistency is a statistical technique for improving answer reliability. Tree of Thoughts is a structured search strategy for problem-solving.

Yes. Ask the same question 3-5 times in separate messages (or separate chat sessions to avoid the model anchoring on its previous answer). Request that it show its reasoning each time. Compare the final answers manually. This is slower than automated sampling but captures the same benefit for high-stakes decisions. For everyday use, 3 runs is sufficient — you rarely need more than that to establish a clear consensus.