Test-Time Compute Scales Better with Parallel Agent Teams
With 100 agents and a 27B model, teams of ten beat independent agents on 10 of 12 tasks.
The method runs several small collaborating teams in parallel, then uses a verifier to select the best result. Tests covered long-horizon tasks including GPU kernels, formal proofs and exploit development. It also beat a single 100-agent team on all four tasks where that comparison was run, and a frontier model on half the tasks at equal dollar cost. The authors’ statistical model attributes the advantage to combining stronger average attempts with multiple chances to find an exceptional result.
ArXiv · AI/CL/LG's note
The method runs several small collaborating teams in parallel, then uses a verifier to select the best result. Tests covered long-horizon tasks including GPU kernels, formal proofs and exploit development. It also beat a single 100-agent team on all four tasks where that comparison was run, and a frontier model on half the tasks at equal dollar cost. The authors’ statistical model attributes the advantage to combining stronger average attempts with multiple chances to find an exceptional result.
ArXiv · AI/CL/LG's note
score 6