Voting / Self-Consistency

What Problem It Solves

For many prompts, the model is stochastic. Voting reduces variance by sampling multiple answers and selecting a winner.

When to Use

• Answers are short and easy to normalize.
• The task is cheap enough to sample N times.
• You want robustness more than latency.

Core Flow

flowchart TD
  P["Same prompt"] --> S1["Sample 1"]
  P --> S2["Sample 2"]
  P --> S3["Sample 3"]
  S1 --> V["Vote/majority"]
  S2 --> V
  S3 --> V
  V --> O["Winner"]

Evolution Path

• Often paired with: Maker-Checker, CoVe (verify claims after voting)
• In production: add evals to detect regressions

Repo Reference

• Code: src/agent_patterns_lab/patterns/voting.py
• Example: examples/31_voting.py
• Tests: tests/test_voting.py