Self-Consistency (CoT-SC) Pattern

Sample multiple reasoning paths and select the most consistent answer. Achieved +17.9% on GSM8K, +12.2% on AQuA.

Activation

Use for:

• Problems with fixed answer sets (math, multiple choice)
• When single-path reasoning is unreliable
• High-stakes decisions requiring confidence
• Factual queries where accuracy is critical

Process

1. Sample Diverse Paths

Generate N (typically 5-10) reasoning paths with higher temperature.

2. Extract Final Answers

Parse the conclusive answer from each path.

3. Majority Vote

Select the most frequent answer across all paths.

Implementation

python
def self_consistency(prompt, question, num_samples=5):
    answers = []
    for _ in range(num_samples):
        response = reason_with_cot(prompt, question)
        answer = extract_final_answer(response)
        answers.append(answer)
    
    # Majority vote
    from collections import Counter
    return Counter(answers).most_common(1)[0][0]

When NOT to Use

• Free-form generation (no fixed answers)
• Time-critical tasks (adds latency from multiple samples)
• Simple queries (overkill)
• Creative tasks (want diversity, not consensus)

Universal Self-Consistency

For open-ended tasks without fixed answers:

"Here are several responses to the same question:
[Response 1]
[Response 2]
[Response 3]

Determine which elements are most consistent across these responses
and synthesize the most reliable answer."

Integration

• Use with @test-automator for test case generation
• Combine with reflection for verified answers
• Log consistency scores to learning engine

Based on Google Research - arXiv:2203.11171