Parallel Sampling

ParallelSamplingStepDescription runs the same reasoning N times and aggregates the candidates — majority vote or an LLM judge. This is the “LLM council” pattern.

from mmar_carl import (
    ParallelSamplingStepDescription, ParallelSamplingStepConfig, ParallelSamplingAggregation,
)

ParallelSamplingStepDescription(
    number=1,
    title="Sample answers",
    aim="Answer the question.",
    config=ParallelSamplingStepConfig(
        n_samples=5,
        aggregation=ParallelSamplingAggregation.MAJORITY_VOTE,
    ),
)

ParallelSamplingStepConfig

Field	Default	Purpose
n_samples	—	How many candidates to sample.
aggregation	—	How to pick a winner (below).
judge_prompt	""	Prompt for the best_of_n / llm_judge judge.
normalize_for_vote	—	Normalise text before majority comparison.

Aggregation strategies

ParallelSamplingAggregation	Picks the winner by…
MAJORITY_VOTE	most common response (exact / normalised match).
BEST_OF_N	an LLM judge selects the best candidate.
LLM_JUDGE	alias for best_of_n with an explicit judge_prompt.

Tip

Run the council across different models by combining sampling with per-step llm_config — see the LLM council example.

See also

• Debate — structured argument instead of independent samples.