self-consistency

**Self-consistency** is the **reasoning strategy that samples multiple independent solution paths and selects the most frequent final answer** - it improves robustness by aggregating over stochastic reasoning variation. **What Is Self-consistency?** - **Definition**: Multi-sample inference method where the same prompt is run several times with non-zero randomness. - **Aggregation Rule**: Final output chosen by majority or highest-consensus answer among sampled paths. - **Use Context**: Primarily applied to reasoning-heavy tasks with one objectively correct target. - **Compute Cost**: Requires multiple model calls, increasing latency and inference expense. **Why Self-consistency Matters** - **Accuracy Gain**: Consensus often filters out unstable single-sample reasoning errors. - **Robustness Improvement**: Reduces sensitivity to one unlucky decoding trajectory. - **Confidence Signal**: Agreement rate can serve as a practical uncertainty indicator. - **Method Compatibility**: Works well with chain-of-thought and decomposition approaches. - **Production Tradeoff**: Benefits must be balanced against throughput and cost constraints. **How It Is Used in Practice** - **Sampling Policy**: Choose sample count and temperature based on quality target and budget. - **Answer Normalization**: Standardize equivalent outputs before voting. - **Fallback Logic**: Escalate low-consensus cases to stronger models or human review. Self-consistency is **a practical ensemble-style inference method for reasoning tasks** - majority aggregation across multiple paths frequently delivers more reliable final answers than single-pass decoding.