Self-Consistency Improves Chain of Thought Reasoning in Language Models

Chain-of-thought (CoT) prompting enables language models to tackle complex reasoning by generating intermediate steps before arriving at a final answer. Typically, this process relies on greedy decoding—simply picking the most likely next token at every step. This study introduces a more robust alternative called self-consistency, which replaces greedy decoding with a sample-and-marginalize strategy. Instead of relying on a single chain of thought, the model generates multiple diverse reasoning paths and selects the answer that appears most frequently among them. This approach is based on the intuition that even if a model makes mistakes, the correct reasoning path is more likely to yield a consistent answer across different samples than an incorrect one.

Self-consistency works by sampling multiple outputs from the model's decoder for a single prompt. For each sample, the model generates a rationale and a final answer. Once a set of these outputs is collected, the method applies a majority vote (marginalization) over the final answers to determine the most consistent prediction. Importantly, this method is unsupervised and requires no additional training, fine-tuning, or external verifiers. It acts as a form of self-ensemble, leveraging the model's existing probabilistic space to improve performance without requiring extra human annotations or auxiliary tasks.

Across a wide array of arithmetic and commonsense reasoning benchmarks—including GSM8K, SVAMP, and StrategyQA—self-consistency consistently outperformed standard greedy CoT prompting. The gains were particularly striking for larger models like PaLM-540B and GPT-3. For example, on the GSM8K arithmetic benchmark, self-consistency achieved accuracy improvements of up to 17.9% compared to traditional greedy CoT. The study also demonstrates that self-consistency is robust across different sampling strategies and can even improve performance in scenarios where standard CoT prompting would normally decrease accuracy.