v8 · architecture
Self-Consistency
sample in parallel, judge the best
Self-Consistency architecture
v8 self-consistencyThe same task is sampled several times independently and in parallel; a judge then either selects the strongest answer or merges the candidates into a consensus. Best for questions where a single attempt is unreliable but agreement across attempts signals a good answer.
Sampler
Judge
sample in parallel, judge the best
How it works
- 1The task is sent to N sampler agents that run independently and in parallel.
- 2Each produces its own complete answer, unaware of the others.
- 3A judge reads all N candidates.
- 4The judge selects the strongest, or merges complementary answers into one.
When to use
Questions where one attempt is noisy but agreement across attempts signals quality.
Trade-off
N parallel samples cost N× the tokens of a single attempt for the sampling step.
Agents
SamplerProduces one independent attempt (×N in parallel)
JudgeSelects the best sample or merges a consensus
Note
Cheap reliability boost on reasoning-heavy questions: sample a few times, let a judge reconcile. The judge’s merge step is where a lot of the value is — it’s not just a vote.