What good is running 64 parallel AI attempts if the step that combines them was never trained to do it?

This explores whether the payoff from sampling many parallel reasoning traces gets thrown away when the aggregator combining them was never trained for the task. The corpus's sharpest answer is yes, and it comes from a clean decomposition of test-time compute into three primitives: sequential (a single long chain), parallel (many independent samples), and aggregative (merging those samples into one answer). Post-training pipelines optimize only the first — models learn to reason step-by-step but never learn to produce jointly informative traces or to arbitrate among them, so parallel compute gets spent generating samples the model has no trained skill for exploiting Why does post-training ignore parallel and aggregative reasoning?. The benefit of drawing 64 samples is real; the machinery to cash it in is missing.

What makes an aggregator good turns out to depend on what it's fed, which flips the usual intuition. Set-level reinforcement learning rewards a batch of traces for being *collectively* useful rather than individually correct, and the surprising result is that diverse-but-mediocre traces beat redundant-but-strong ones — because an aggregator needs raw material to arbitrate between, not sixty copies of the same confident guess Can diverse mediocre traces outperform redundant expert traces?. An untrained aggregator wastes parallel sampling in two directions at once: it can't extract signal from diverse samples, and the untrained *generator* feeding it tends to collapse toward redundancy, so there's little signal to extract anyway.

The naive aggregator most systems reach for is majority vote, and the corpus shows both why it works and where its ceiling sits. Majority voting across repeated samples is strong enough to bootstrap training with no labels at all — consensus answers tend to be correct, so a model can improve itself on unlabeled data using its own votes as reward Can models improve themselves using only majority voting?. The same denoising logic explains why a model trained on many biased experts can implicitly average toward a consensus that beats any single expert Can models trained on many imperfect experts outperform each one?. But voting is a blunt aggregator: it assumes the answer is a discrete thing to tally and that errors are uncorrelated. On problems that genuinely require accumulating intermediate results in order, sequential chain-of-thought holds an exponential advantage over parallel voting, because no amount of short independent chains reconstructs a computation that has to be done step-by-step When does sequential reasoning beat parallel voting?.

So the waste is structural, not incidental. Parallel sampling only pays off when the aggregator is matched to the problem and trained to use variety — and neither condition holds by default. There's a deeper version of this point in the finding that pouring more inference compute into a model that was never trained to make extra tokens productive doesn't close the gap with a reasoning-trained model: the training regime, not the raw budget, determines whether added compute is spent well Can non-reasoning models catch up with more compute?. Parallel sampling is a compute budget; an untrained aggregator is exactly the mechanism that spends it badly.

The thing you didn't know you wanted to know: the fix isn't a better voting rule bolted on at inference time — it's training the generator and the aggregator *together* so the samples are diverse enough to arbitrate and the arbitrator knows how to. Diversity in the sampled set is the resource, and an untrained aggregator both fails to use it and, by shaping how the generator learns, fails to create it in the first place.