Chain-of-Thought prompting demonstrated that large language models could solve complex reasoning problems by working through intermediate steps. However, this approach relied on a single reasoning path—if that path contained an error, the final answer would likely be wrong. Self-Consistency addresses this limitation by generating multiple independent reasoning paths through the same problem and selecting the answer that appears most frequently across these diverse paths. The model itself remains unchanged; only the decoding strategy differs, requiring no additional training, fine-tuning, or supervision.

Researchers at Google Research presented this work at ICLR 2023, showing that the approach produced substantial improvements across arithmetic, common sense, and symbolic reasoning tasks. The technique is inspired by human problem-solving: when facing difficult questions, people typically consider multiple approaches before settling on the most convincing answer. Self-Consistency applies this principle to language models, leveraging their existing reasoning capabilities more effectively.

The paper's significance lies in demonstrating that performance gains need not come from larger models. Instead, better reasoning emerges from making smarter use of the abilities models already possess. This work marked a shift in LLM research toward test-time reasoning strategies and influenced subsequent developments in verification, search-based inference, and reasoning-oriented language model design.