Test-Time Compute Scales Better with Parallel Agent Teams
The paper proposes parallel agent teams, a test-time scaling method that samples many small collaborating agent teams in parallel and keeps the best team's result. Evaluated on 12 long-horizon tasks spanning GPU kernels, formal proofs and exploit development, teams of ten with a 27B model and a 100-agent budget beat parallel independent agents on 10 of 12 tasks and beat a single team of 100 on all four tasks where such large teams were run. At equal dollar cost, the parallel teams beat a frontier model on half the tasks. A statistical model attributes the gains to balancing a higher mean with variance-based selection benefits.
It offers a concrete, empirically tested strategy for spending inference compute more effectively with multi-agent teams, relevant to practitioners scaling agent systems.
Comments & Questions (0)
No comments yet
Be the first to comment!