Self-Consistency Improves Chain of Thought Reasoning in Language Models

1 Introduction

1. What does self-consistency replace in chain-of-thought prompting?
   The greedy decoding of a single reasoning path. Self-consistency samples many diverse reasoning paths from the decoder and returns the answer most of them agree on.
   • Wang et al., 2023, Self-Consistency Improves Chain of Thought Reasoning in Language Models (The Eleventh International Conference on Learning Representations (ICLR 2023)), §1
2. What intuition motivates self-consistency?
   A complex reasoning problem usually admits several different ways of thinking that lead to the same correct answer, while wrong reasoning paths tend to scatter over different wrong answers; if several independent lines of thought agree, the answer deserves more confidence.
   • Wang et al., 2023, Self-Consistency Improves Chain of Thought Reasoning in Language Models (The Eleventh International Conference on Learning Representations (ICLR 2023)), §1
3. Which problems of single-sample decoding does self-consistency avoid?
   The repetitiveness and local optimality of greedy decoding, and the randomness of relying on one sampled generation.
   • Wang et al., 2023, Self-Consistency Improves Chain of Thought Reasoning in Language Models (The Eleventh International Conference on Learning Representations (ICLR 2023)), §1
4. How does self-consistency differ from verifiers, re-rankers and ensembles?
   It needs no extra training, verifier, re-ranker, human annotation or fine-tuning, and it works off the shelf with one pretrained model, acting as a “self-ensemble” rather than aggregating several trained models.
   • Wang et al., 2023, Self-Consistency Improves Chain of Thought Reasoning in Language Models (The Eleventh International Conference on Learning Representations (ICLR 2023)), §1