The authors apply Steiner's taxonomy of group tasks to multi-agent LLM systems, separating disjunctive tasks with a single best answer from compensatory tasks where continuous estimates are aggregated. They model independently sampled agents as conditionally independent given the item, yielding large-team limits: plurality voting converges to the model's modal answer, and averaging converges to the model's item-level bias. Experiments use 13 open-weight models, teams of up to 30 agents, and representative benchmarks for both task types.
On disjunctive tasks, the chance that at least one agent is correct rises by 5 to 20 percentage points with team size, but plurality voting captures almost none of this gain, matching predictions within 0.5 points. Multi-round revision raises accuracy considerably, yet one peer is nearly as good as 29. On Fermi estimation, averaging reduces error by only about 6% because shared item-level biases account for about 87% of squared error. Combining model families helps Fermi estimation but does not surpass the strongest member on disjunctive tasks.
This is interesting because it shows that simply adding agents is not a universal path to better performance. The task structure and the mechanism for combining outputs, not just team size, determine whether scaling helps.