Self-Consistency Improves Chain of Thought…

A model can reach the same answer along several different routes. This paper shows how sampling those routes and voting on their final answers improves reasoning accuracy without additional training, then exposes the extra computation and the gap between agreement and correctness.

Reading focus: How self-consistency replaces one greedy chain of thought with several sampled solutions and a vote over final answers. Why grouping answers differs from selecting the most probable individual explanation. How to read the accuracy gains alongside generation cost, answer extraction, and repeated mistakes.

ICLR 2023. Xuezhi Wang et al.. 28 min read, easy difficulty.