When solving a hard problem with a language model, running it multiple times and picking the best result works better than one try. However, many separate tries give limited gains because each try alone is not very capable. Making many copies work together as one team can produce a stronger result, but you get only one outcome, losing the benefit of trying different approaches.
A new method, parallel agent teams, runs many small teams of collaborating models in parallel and keeps the best result. Each small team is more effective than a single model, and having many teams provides diverse solutions.
On real-world tasks like writing fast code or proving theorems, parallel teams beat both separate runs and one large team, achieving up to 27% better results and continuing to improve with more compute. This offers a simple way to get more from the same model.