Self-consistency improves reasoning accuracy by generating multiple solution paths and selecting the most common answer. Mechanism: Sample N reasoning chains with temperature > 0, extract final answer from each chain, return majority answer (modal response). Why it works: Correct reasoning paths more likely to converge on same answer, errors tend to be random/diverse, voting filters out inconsistent mistakes. Implementation: Generate 5-40 chains, parse answers (often needs structured output), count occurrences, return mode. Cost trade-off: N× more expensive than single chain, but significantly higher accuracy on complex reasoning. When to use: Math problems, logical reasoning, factual questions with objective answers, high-stakes decisions. Limitations: Doesn't help if model is systematically wrong, expensive for production, requires parseable answers. Optimal N: 5-10 often sufficient, diminishing returns beyond 20. Variants: Weighted voting by confidence scores, minimum consistency threshold before answering, combining with ToT exploration. Results: 10-20% accuracy improvements on benchmarks like GSM8K, significant for mathematical reasoning.