Self-consistency improves chain-of-thought prompting by generating several independent reasoning traces for the same question, extracting each trace's final answer, and selecting the most common one. The intuition is that correct reasoning paths are more likely to converge on the same answer, while errors are more random. It trades compute cost for accuracy and is especially effective on math and logic problems.