What they found
The authors compare three ways to spend extra computation at test time when a verifier can score each attempt. One is many independent agents, keeping the best (best-of-N). Another is one large team of collaborating agents. Their method, parallel agent teams, runs many small teams at once and keeps the best team's result. With a 27B model and a budget of 100 agents, parallel teams of ten beat parallel agents on 10 of 12 tasks. They also beat one team of 100 on all four tasks where such large teams were run. At equal dollar cost, the same parallel teams beat a frontier model on half the tasks. The tasks were long-horizon ones, from GPU kernels and formal proofs to exploit development. A simple statistical model explains the result. Solo agents have a low average score. One big team gets only one sample, so it gains nothing from variance.
Why it matters
Spending more computation at inference is a main way to get more from a fixed model. This paper suggests how to split that spending between teamwork and repeated tries. The result applies when a verifier can score every attempt. The abstract does not say how well it holds without one.
How it changes the race
If it holds, a smaller, cheaper model run as many small teams could match or beat a frontier model on some hard tasks at the same cost. That would help groups without the largest models. The abstract reports this on half the tasks, not all.
Safety and control
One of the 12 tasks is exploit development, so this kind of scaling may also raise capability in security-relevant work. The abstract does not discuss safeguards or misuse.
The abstract, by the authors
Test-time scaling improves the performance of a fixed model by spending more computation at inference. There are two common ways to spend it on agents when a verifier can score every attempt. Best-of-N sampling runs many independent agents and keeps the best result but saturates quickly. One large team of collaborating agents can score higher but stops improving as it grows. Can we get the best of both? Our method, parallel agent teams, does so by sampling many small teams in parallel and keeping the best team's result. We test it on 12 long-horizon tasks from GPU kernels and formal proofs to exploit development. With a 27B model and a budget of 100 agents, parallel teams of ten beat parallel agents on 10 of 12 tasks and beat one team of 100 on all four tasks where we ran such large teams. At equal dollar cost the same parallel teams beat a frontier model on half the tasks. A simple statistical model shows that individual agents lose because their many samples have a low mean and that one large team loses because a single sample gets no gain from variance. Sampling many small teams in parallel balances a high mean with the gain from variance. Parallel agent teams thus offer a more effective way to scale test-time compute.
Quoted from arXiv:2611.09798. The research and its words belong to its authors.