Self-Consistency Prompting
A decoding strategy that samples multiple reasoning paths and selects the most consistent answer through majority voting.
Also known as: Self-Consistency Decoding, Majority Voting Prompting, CoT-SC
Category: Techniques
Tags: ai, prompting, reasoning, llm-techniques, reliability, ensemble-methods
Explanation
Self-consistency prompting improves Chain-of-Thought reasoning by sampling multiple diverse reasoning paths and selecting the answer that appears most frequently. Instead of relying on a single greedy decode, it leverages the intuition that correct reasoning paths are more likely to converge on the same answer.
How it works:
1. **Generate multiple paths**: Use temperature sampling to generate several different reasoning chains for the same problem
2. **Extract answers**: Parse the final answer from each reasoning path
3. **Majority vote**: Select the answer that appears most frequently across all paths
Why it works:
- Complex problems often have multiple valid reasoning approaches
- Errors in reasoning tend to be random and don't consistently lead to the same wrong answer
- Correct approaches naturally converge on the true answer
Implementation considerations:
- **Sample count**: Typically 5-40 samples; more samples generally improve accuracy but increase cost
- **Temperature**: Higher temperature (0.5-1.0) encourages diverse reasoning paths
- **Answer extraction**: Requires reliable parsing to extract comparable answers
Self-consistency is particularly effective for:
- Mathematical reasoning
- Commonsense reasoning
- Tasks with verifiable answers
- Problems where multiple solution approaches exist
This technique consistently outperforms single-path Chain-of-Thought across various benchmarks, trading computation for accuracy.
Related Concepts
← Back to all concepts