| name | self-consistency |
| description | Generates multiple independent reasoning paths and selects the answer by majority vote or consensus. Use for verification, catching edge cases, high-stakes decisions, or any problem where multiple valid approaches exist. Triggers on "verify", "double-check", "are you sure", or when confidence in a single approach is insufficient. |
Self-Consistency
Samples multiple reasoning paths and aggregates for reliability.
Core Mechanism
Instead of one reasoning chain, generate several independent chains and vote:
Problem → Chain A → Answer A
→ Chain B → Answer B
→ Chain C → Answer C
→ Majority vote → Final Answer
Process
1. Generate multiple diverse reasoning paths (3-5 typically)
2. Ensure independence: each path should approach differently
3. Extract the answer from each path
4. Aggregate:
- Majority vote for categorical answers
- Consensus check for open-ended responses
- Flag significant disagreements for review
Key Principles
- Independence: Paths should not influence each other
- Diversity: Different approaches, not just different wordings
- Aggregation: The wisdom of crowds, but for reasoning
- Discrepancy detection: Disagreements surface edge cases or errors
When to Apply
- High-stakes decisions where errors are costly
- Problems with multiple valid solution approaches
- Verification of important conclusions
- When single-path confidence is insufficient
- Edge case detection
Implementation Patterns
Simple Multi-Path
Solve this problem three different ways:
Approach 1 (algebraic):
[reasoning → answer]
Approach 2 (geometric/visual):
[reasoning → answer]
Approach 3 (estimation/sanity check):
[reasoning → answer]
Consensus check: Do all approaches agree? If not, identify the discrepancy.
Implementation Verification
Generate two implementations:
Implementation A (optimize for clarity):
[code]
Implementation B (optimize for efficiency):
[code]
Verification:
- Do they produce the same output for test cases?
- If different, which is correct and why?
Decision Verification
Evaluate this decision from three perspectives:
Perspective 1 (short-term impact):
[analysis → recommendation]
Perspective 2 (long-term impact):
[analysis → recommendation]
Perspective 3 (risk assessment):
[analysis → recommendation]
Synthesis: What does the consensus suggest?
Handling Disagreements
When paths disagree:
- Identify where the reasoning diverges
- Check each path for errors at the divergence point
- If both seem valid, the problem may be ambiguous
- Report the disagreement and reasoning for each
Anti-Patterns
- Paths that aren't actually independent (just rephrased)
- Ignoring disagreements (they're the valuable signal)
- Too many paths (diminishing returns after 5)
- Using for simple problems where one path suffices