Parallel-sample
The parallel-sample loop answers a question K times independently, then reconciles the K answers into one. Because the samples don’t see each other, their agreement is signal: when independent attempts converge, the consensus is more trustworthy than any single attempt. Use it when a question has a checkable answer and you’d rather spend compute on diversity than on a longer single chain.
This page explains the loop. For how a loop is chosen, see Coordination loops.
The loop
Section titled “The loop”Parallel-sample is its own skeleton — a fan-out-join-reduce topology, not a variant of VMAO. There are no plan, verify, or replan phases to counterfeit, so it is a first-class coordinator over the same governed primitives every loop shares:
- SAMPLE — one governed wave of K lanes, all dispatched together. The wave passes through the shared dispatch path (per-wave budget carve, bounded concurrency, per-member resume), exactly as VMAO’s execute waves do. K is a configuration value; the loop is bounded by construction, so it carries no iteration caps.
- Barrier — a checkpoint joins the completed samples. Loop progress rides the event log, so an operator watches the samples land in real time.
- REDUCE — the K answers collapse to one, two ways (see below), producing a result-bearing outcome.
Two reducers, one skeleton
Section titled “Two reducers, one skeleton”The reducer is a configuration choice, not a different loop:
- Majority vote — a pure fold over the samples’ extracted answers; the most agreed-upon answer wins. This is self-consistency.
- Judge + argmax — a second wave dispatches a judge that scores each sample, and the highest-scoring one wins. This is best-of-N.
Both are two configurations of the same fan-out-join-reduce skeleton; the vote fold and the judge-scoring parse are shared substrate primitives, so the two reducers never diverge.
Research lineage
Section titled “Research lineage”Majority-vote sampling is self-consistency (Wang et al., 2022); verifier-argmax is best-of-N. The plane’s economics are the point: several of the field’s named strategies are configurations of one skeleton — fan-out, barrier, and a reducer — not separate implementations.
When to choose it
Section titled “When to choose it”Choose parallel-sample when a question has a checkable or votable answer and independent attempts add confidence: a factual answer where consensus matters, a generation task where you want the best of several tries. If the work needs verification and revision of decomposed sub-tasks, use VMAO; if you want the samples to see and rebut each other rather than stay independent, use debate.
Where this shows up
Section titled “Where this shows up”- Reference: the orchestrator’s parallel-sample coordinator under Reference.