Self-Consistency: Improving Chain-of-Thought Reasoning

07/10/26 on LLM Papers

Paper: Wang, X., et al., Self-Consistency Improves Chain-of-Thought Reasoning in Language Models, ICLR 2023. arXiv:2203.11171

Why this paper matters

Chain-of-Thought (CoT) showed that asking a model to “think step by step” unlocks reasoning. But a single greedy reasoning path is fragile — one bad step derails the answer. Self-Consistency makes CoT robust with a tiny change: sample many reasoning paths and take the majority vote on the final answer. It’s a decoding trick, not a training trick, and it lifted accuracy on math and commonsense benchmarks by double digits. If you do any reasoning with LLMs, this is the cheapest gain you can get.

The core idea: marginalize over reasoning paths

A reasoning problem usually has many valid chains that reach the same answer but few that reach a wrong one. Greedily picking one path ignores that. Instead:

Sample $k$ diverse CoT paths ${z_1, \dots, z_k}$ with temperature $> 0$.
Parse the final answer $a_i$ from each path.
Aggregate by majority vote (or weighted by path probability):

\[\hat{a} = \operatorname*{argmax}_{a} \sum_{i:\, \text{ans}(z_i) = a} p(z_i \mid x)\]

Intuition: the correct answer is a fixed point — many independent samplings converge to it; wrong answers scatter.

A minimal sketch

paths = [model.generate(prompt, temperature=0.7) for _ in range(20)]
answers = [parse_answer(p) for p in paths]
final = max(set(answers), key=answers.count)   # majority vote

No extra training, no verifier model required.

Key results

GSM8K: ~70% → ~78% (and higher with more samples) for PaLM 540B.
Large gains on MATH, StrategyQA, ARC-challenge, and other multi-step tasks.
Near or surpassing supervised verifier methods at the time, with zero additional supervision.

Why it matters today

Self-Consistency is the practical companion to the CoT post. CoT gives the model a reasoning process; self-consistency makes that process reliable. It’s also the conceptual ancestor of “test-time compute / sampling many thoughts” ideas that reasoning models (and Tree-of-Thought search) build on.

References

Wang et al. (2022). Self-Consistency Improves Chain-of-Thought Reasoning in Language Models. ICLR 2023. arXiv:2203.11171
Wei et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arXiv:2201.11903