Glossary term
Glossary term
Architecture
Sampling multiple reasoning chains and selecting the most consistent answer.
Wang et al. (2022) showed that sampling 40 CoT reasoning chains from PaLM 540B and majority-voting the answers improves GSM8K accuracy from 56.5% to 74.4% - a significant gain with no additional training.
Perplexity AI uses self-consistency-like answer aggregation across multiple search queries and model responses - combining diverse retrieved results and generation samples to produce more robust factual answers.
A medical AI diagnostic system uses self-consistency: generating 10 diagnostic reasoning chains and selecting the diagnosis supported by 7+ chains - reducing single-chain reasoning errors by 23% in clinical trials.