Chain of thought and self-consistency: majority voting, and when it fails
Why writing intermediate steps helps, how sampling several chains and voting improves accuracy, a runnable simulation showing the gain (40% to 75% at 21 samples) and the failure when errors are correlated (to 18%), and the cost arithmetic.
18 min·Verified 2026-11Pro+50 XP
The simplest way to buy accuracy with compute is to let the model write out its reasoning, then ask more than once.
Pro content
Keep reading with Inference Academy Pro
Unlock the rest of this lesson, plus every advanced lesson, hands-on lab, the larger playground models and completion certificates.