Self-Consistency Multi-Path Prompt

What this prompt does

Samples N independent reasoning chains for the same problem, then aggregates them using majority vote (or a second aggregation call) to arrive at the most reliable answer. Reduces the impact of any single incorrect reasoning path.

When to use it

• High-stakes reasoning tasks where a single wrong chain causes failures (math, logic, medical triage)
• When you cannot do few-shot calibration but need higher accuracy than zero-shot
• Offline evaluation pipelines where latency is less important than accuracy

When NOT to use it

• Real-time user-facing applications — N calls multiplies latency and cost
• Tasks with a single clearly correct format (extraction, classification) — simpler methods suffice

The Prompt

Phase 1: Generate N independent paths (run N times with temperature > 0)

Phase 2: Aggregate (run once with the N outputs)

You are given {{N}} independent solutions to the same problem. Your job is to identify the most consistent answer.

{{SOLUTION_1}}
---
{{SOLUTION_2}}
---
{{SOLUTION_3}}

Instructions:
1. List the final answer from each solution.
2. Identify which answer appears most frequently.
3. If there is a tie, pick the answer supported by the most rigorous reasoning.

Final Answer: [The most consistent answer]
Confidence: [HIGH if majority agree, MEDIUM if split, LOW if all differ]

Why it's structured this way

"Do not look at any previous attempts" — Prevents the model from anchoring to its own prior output when run in a multi-turn context. Forces true independence between paths.

Temperature > 0 for Phase 1 — Diversity in reasoning paths is the mechanism. Running at temperature 0 produces near-identical paths and defeats the purpose.

Explicit aggregation in Phase 2 — Naive string matching on "Final Answer:" is fragile. A second LLM call handles paraphrased answers (e.g., "42" vs "forty-two") better than regex.

Confidence field — Surfaces when the model is genuinely uncertain (all paths differ), enabling your application to escalate to a human reviewer.

Variations