Self-Consistency — Voting Across Multiple Reasoning Paths

Why it matters

Self-consistency is a cheap way to boost reasoning reliability. No new training, just extra inference. For high-stakes tasks, the extra tokens are worth it.

The architecture

Sample N different CoT completions using temperature > 0. Each produces its own reasoning trace and final answer.

Aggregate: take the most common final answer. Optionally return confidence based on vote fraction.

How it works end to end

Sampling diversity: temperature 0.7-1.0 gives useful diversity. Too low: all traces identical. Too high: garbage traces.

N selection: 5-10 traces balance cost and reliability. Beyond 10, diminishing returns.

Weighted voting: some techniques weight votes by trace confidence or quality score.