Definition
Plain language
Combining several separate attempts or answers into one final answer, hoping the group is better than any individual.
As stated in the literature
Aggregating predictions from multiple models or runs (voting, averaging, stacking) to reduce variance; costs proportional extra inference calls, which is why the evolved agent discarded it under a fixed budget.
Also called: ensemble, ensembles
Why it matters: It reliably smooths out individual mistakes, but it multiplies compute cost, so under a fixed budget it can be the first thing worth cutting.
For example, instead of trusting one model's guess at a house price, you ask five models and average their answers.
Heard on the show
“So search keeps spending hours rediscovering "math wants ensemble voting," and then throws that knowledge away before the next math task starts.”Episode 013 — Why Search Keeps Rediscovering the Same Workflow, and What That Means