Published: 2022

Self-Consistency Improves Chain of Thought Reasoning in Language Models

CATEGORIES

RISK-BASED PROCESS SAFETY ELEMENTS

Research Summary

Introduces self‑consistency for chain‑of‑thought prompting: instead of taking one greedy reasoning trace, sample multiple reasoning paths and select the most consistent final answer across those samples. This ensemble-style decoding improves correctness on several reasoning benchmarks and reduces sensitivity to minor prompt changes. In safety‑critical PSM contexts, output stability matters—two runs should not produce wildly different initiating events, safeguards, or recommendations. Self‑consistency suggests a practical reliability control: generate multiple independent hazard analyses or compliance checks, compute consensus, and flag low‑agreement items for human review. That supports defensible governance around LLM-assisted PSM work products.

AUTHORS

Xuezhi Wang; Jason Wei; Dale Schuurmans; Quoc V. Le; Ed H. Chi; Sharan Narang; Aakanksha Chowdhery; Denny Zhou

CITATIONS

X. Wang et al., "Self-Consistency Improves Chain of Thought Reasoning in Language Models," arXiv:2203.11171, Mar. 2022.