Self-Consistency Improves Chain of Thought Reasoning in Language Models

ICLR 2023 Google Research, Brain Team arXiv 2203.11171

TL;DR

Self-Consistency boosts the performance of chain-of-thought prompting with a striking margin in a “self-ensemble” manner.

Motivations & Innovations

• Although LLMs have demonstrated remarkable success, their reasoning capabilities are still limited, which cannot be overcome solely by increasing model scale. -> chain-of-thought prompting
• LLMs are not perfect resoners -> incorrect reasoning paths -> self-consistency to boost the performance

Approach

Self-Consistency over Diverse Reasoning Paths

• Few-shot Chain-of-Thought (CoT) Prompting
• Sample a diverse set of candidate reasoning paths instead of only taking the greedy one
  • temperature sampling
  • top-k sampling
  • nucleus sampling / top-p sampling
• Marginalize out the sampled reasoning paths by taking a magority vote

Experiments

• self-consistency boosts the performance of chain-of-thought prompting with a striking margin

Robustness of Self-Consistency over Diverse Sampling Strategies: Self-Consistency is robust to sampling strategies and scaling.