Self-Consistency & Voting

Voting

Part of: Advanced Prompting Patterns

A model solves a tricky word problem and answers 42. Run the same prompt again and it answers 27. Run it five more times: 42, 42, 27, 42, 31. Which do you trust? Count them: four of seven say 42, and that majority answer is more reliable than any single run. This is self-consistency , and it can lift accuracy several points for the cost of running the prompt a few more times. What it is Self-consistency is a pattern where you sample multiple chain-of-thought answers to the same question, then take the majority vote of the final answers. Instead of trusting one reasoning path, you generate several diverse paths and let them vote. It works because a hard problem has many wrong answers but usually one right one. Correct reasoning tends to converge on the same final answer through different routes, while errors scatter. Sample enough paths and the correct answer clusters at the top while mistakes spread thin. How it works The recipe is short: 1. Set a non-zero temperature so each run takes a slightly different reasoning path. At temperature zero every run is identical and voting is pointless. 2. Sample N reasoning chains for the same prompt (say N = 5). 3. Extract each final answer fro

Challenge: Confidence-Weighted Vote