Self-Consistency

Overview

Proposed by Google in 2022, Self-Consistency samples multiple (typically 10-40) CoT reasoning paths with temperature > 0 and takes a majority vote on the final answers. It substantially improves accuracy on arithmetic, commonsense reasoning, and code generation compared with single greedy decoding.

Cost tradeoffs

Multiple inference calls increase API cost and latency linearly. Self-Consistency suits accuracy-critical, latency-tolerant scenarios (weekly report generation, batch processing), not real-time chat.