Neura saysSelf-consistency samples several reasoning paths and takes the majority answer.
Instead of trusting one chain-of-thought, self-consistency samples multiple independent reasoning paths (with some randomness) and picks the most common final answer. Different paths that converge on the same answer are more likely correct; outliers are filtered out. It reliably boosts accuracy on reasoning tasks — at the cost of running the model several times. A simple, effective inference-time technique.
Power-ups you unlock
Sample multiple reasoning paths
Take the majority final answer
Convergent answers are more reliable
Costs several model runs
The Overfit Ogre attacks — common mistakes
Using one sample for hard reasoning
Majority-voting on tasks with no single answer
Ignoring the added cost
Boss battleExplain why majority-voting over sampled paths beats a single path.