← Concept Index

Reasoning & inference

Self-consistency and verification

Also called: double-checking answers, sampling multiple solutions

DEFINITION

Techniques that improve reliability by having a model solve a problem more than once, or check its own answer, and preferring the result that holds up — rather than trusting a single pass.

WHY IT MATTERS

A single generation can go wrong in one bad step. Cross-checking multiple attempts, or asking for an explicit re-check, catches some of those errors — cheaply improving accuracy on reasoning tasks.

COMMONLY CONFUSED WITH

A guarantee of correctness. A model can be consistently wrong, and it can vouch for its own bad answer; verification reduces error, it doesn't eliminate it or replace an external check.