Duke and Princeton paper shows small agent teams beat solo scaling
The non-obvious insight is that diversity across teams matters more than depth within one team — a parallel small-team structure dominates both axis of scaling at the same cost. Engineers building verification-scored agent workflows can apply the recipe directly with no new infrastructure, just a partitioning change. Compared with standard best-of-N with a verifier, which plateaus quickly, this topology offers a concrete way to squeeze more accuracy from the same token budget.
A new paper from Duke and Princeton researchers finds that, at equal compute, many small agent teams run in parallel outperform both solo-agent best-of-N sampling and one giant team. Using a fixed budget of 100 AI agents, running 10 teams of 10 agents in parallel and keeping the best team's result beat both 100 independent solo agents and a single 100-agent team.
The mechanism splits the agent budget into small teams of 10 that share a discussion forum, so each attempt is strengthened by internal teamwork while multiple teams still preserve the diversity of many tries. The paper notes that the common verifier-based approach of running many solo agents and keeping the best output gains flatten quickly, while one large team improves at first but plateaus as it grows.
For practitioners, the practical rule is straightforward: if a verifier can score your agents' work, divide the compute budget into several small teams that share notes and select the best team's answer. The finding applies to anyone building agent pipelines under fixed inference budgets, suggesting a more efficient allocation than simply scaling team size or sample count.
New Duke and Princeton paper finds that, at equal compute, many small agent teams in parallel beat both solo-agent best-of-N and one giant team.
With the same budget of 100 AI agents, running 10 small teams of 10 agents in parallel and keeping the best team’s result beat both 100 solo agents and one 100-agent team.
With a verifier, teams usually run many solo agents and keep the best, but gains flatten fast. One large team does better at first, then stops improving as it grows.
They split the agent budget into teams of 10 that share a forum, then keep the best team's result. Teamwork raises each attempt's average, and several teams still give many tries.
If a verifier can score your agents' work, split the budget into several small teams that share notes, and keep the best team's answer.
This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include agents, reasoning models, Duke, Princeton.
Source attribution: curated from rohanpaul_ai (@rohanpaul_ai).
