Ensembling

In short: Many samples, aggregated by majority or by consistency.

What is: Self-Consistency votes the majority answer across N CoT paths. Universal Self-Consistency (USC) does the same for free-text answers (the model picks the most consistent). DiVeRSe goes further: diverse prompts × multiple paths + scoring.

When to use: High stakes. USC for textual answers; DiVeRSe for maximum coverage.

When to avoid: Simple or cost-sensitive tasks, all of this costs N× tokens.

💡 Tip: Use the model itself as a judge to pick the most consistent answer when no exact vote is possible.

In short: Ask in several ways and aggregate the answers.

What is: Rephrase the same request in several ways, run them all and aggregate. Reduces the result's sensitivity to specific wording, models sometimes answer differently just because of how you asked.

When to use: When you suspect the result depends too much on how the question was written.

When to avoid: Stable, well-defined tasks where wording doesn't change the result.

💡 Tip: If the paraphrases diverge a lot, the problem is usually ambiguity in the question, worth clarifying first.