← All papers
Self-Consistency Improves Chain of Thought Reasoning in Language Models
Wang, Wei, Schuurmans et al. (Google) · 2023 · ICLR 2023
PromptingRead on arXiv
Improves chain-of-thought prompting by sampling multiple reasoning paths and taking a majority vote over final answers, consistently outperforming single-path greedy decoding.
Key Idea
Instead of generating a single chain-of-thought and trusting that answer, Self-Consistency samples multiple diverse reasoning paths (using temperature > 0) and selects the most frequent final answer via majority voting. The intuition: correct reasoning paths are more likely to converge on the same answer, while incorrect paths tend to diverge.
How It Works
- Prompt the model with chain-of-thought exemplars (same as standard CoT)
- Sample N completions with temperature (e.g., T=0.7), generating N different reasoning chains
- Extract the final answer from each chain
- Majority vote: the answer appearing most frequently across the N samples is selected
- No additional training or fine-tuning required, purely an inference-time technique
- Typically N=5-40 samples; diminishing returns beyond ~20
Why It Matters
- Consistent improvements: +5-18% accuracy over greedy CoT on arithmetic (GSM8K: 78%), commonsense (ARC), and symbolic reasoning benchmarks
- Simple and universal: works with any LLM that supports CoT, no model changes needed
- Pioneered the idea of test-time compute scaling, spending more inference compute for better answers, a precursor to o1-style reasoning
- Showed that LLMs encode multiple valid reasoning strategies and sampling reveals this diversity
- Became a standard technique in production LLM pipelines for high-stakes decisions
Key Takeaways for Interviews
- Self-Consistency = sample N chain-of-thought paths + majority vote on final answer
- It's a free lunch in accuracy at the cost of N× inference compute, classic latency/quality tradeoff
- Works because correct reasoning paths converge while errors are diverse (random)
- Key hyperparameters: temperature (0.5-0.7), number of samples (10-20 is usually enough)
- In system design: use self-consistency for high-value decisions (medical, financial) where latency is acceptable; skip for real-time chat
- Precursor to more sophisticated test-time compute methods: Tree of Thoughts, Best-of-N, o1 reasoning