Self-consistency
A reasoning technique where the model samples multiple chain-of-thought traces for the same problem and selects the most common final answer — cheap accuracy boost on math and logic tasks.
Self-consistency replaces single-greedy decoding with N parallel samples, each generated with non-zero temperature so they explore different reasoning paths. The agent then picks the answer that appears most often across samples.
For math and structured reasoning, self-consistency typically lifts accuracy 5–15 percentage points over greedy chain-of-thought at the cost of N× inference. With reasoning models that already do internal sampling (o3, Claude reasoning, Gemini Thinking), the gains shrink — but self-consistency still helps on out-of-distribution tasks.
In production agent loops, self-consistency is most useful for verifier steps — checking whether a tool result is plausible — rather than every reasoning step.
Frequently asked
How many samples should I use for self-consistency?+
Diminishing returns above 5–10 samples on most benchmarks. The "sweet spot" is 5 samples with temperature 0.5–0.7. Going to 40 samples buys you another 1–2 points and 4× the cost.
Does self-consistency work with reasoning models?+
Less so. Reasoning models do internal exploration during their thinking phase, so external sampling adds less. It still helps on tasks where the model is confidently wrong in the same direction every time.