Mind · Reasoning
SelfConsistency
Sample multiple reasoning chains, select by majority
Definition
Mechanism
A variance-reduction technique where the same query is processed N times independently (with temperature > 0), producing N candidate answers. The final answer is selected by majority Aggregate (Mode). Exploits the intuition that correct reasoning paths are more likely to converge on the same answer than incorrect ones. It aggregates multiple reasoning paths via Aggregate to determine the most robust answer.
Invariants
- Aggregation: Final answer determined by plurality or weighted aggregation
- Independence: Each sample must be generated without knowledge of others
- N >= 3: Minimum samples for meaningful convergence
Risks
Failure modes
Consistent Wrongness
All N samples converge on the same incorrect answer (systematic bias).
Cost Explosion
N samples means N times the compute cost.
Relationships
Builds on
References
Infrastructure
Reduce a set to a summary statistic
Nothing in this library builds on this pattern yet.
Reference