Self-consistency is a test-time ensemble that predates reasoning models and still stacks on top of them — sample N chains, vote on answers; training and inference techniques complement rather than replace each other.
Self-consistency is like asking a student to solve a hard problem five different ways and taking the most common final answer. You do not need a special 'reasoning model' class to do that — any chat model with chain-of-thought prompting can use it. Reasoning models are trained to think longer and better internally, but you can still run self-consistency on top at serve time for extra accuracy on math-like tasks. One does not replace the other; they layer.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Candidates often conflate training a model to reason with sampling many chains at inference. This question tests whether you see them as complementary layers in the test-time compute stack.
Self-consistency arrived years before o-series and DeepSeek-R1. Reasoning models did not retire it — they raised the per-sample baseline that self-consistency ensembles build on.
What self-consistency does
Self-consistency is an inference procedure: prompt the model for chain-of-thought, sample N completions with temperature > 0, parse a final answer from each chain, and take the majority vote (or weighted vote).
It requires no weight updates. Any model that can emit CoT benefits — which is why it predates dedicated reasoning checkpoints.
Gains come from diversity: different stochastic paths stumble differently; correct reasoning often converges on the same final answer while errors scatter. On GSM8K-style tasks, voting lifted standard PaLM accuracy materially.
Cost is linear in N: N forward passes, N times the tokens, N times the latency unless batched.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- Google's self-consistency paper (2022) boosted GSM8K with standard PaLM before o-series reasoning models existed.
- OpenAI o3 and DeepSeek-R1 deployments still expose temperature and multi-completion parameters for best-of-N serving patterns.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow does best-of-N differ from self-consistency?
Self-consistency majority-votes extracted answers; best-of-N scores each full completion with a verifier or RM and picks the top — different aggregation, similar multi-sample cost.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Choosing option A — self-consistency does not make reasoning-model training obsolete; it is an additive inference trick with its own latency and cost tradeoffs.
60 second bullets to scan on the way to the call.
Define self-consistency in one sentence
Explain why it is test-time, not training-time
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.