GPT-5.4 vs MiniMax M3

GPT-5.4 analyzed the simulation's physics, match loop, and player intent files to write a 561-line soccer-team policy. When initial testing against a "gegenpress" opponent resulted in four straight losses and a -6.0 average goal differential, the model modified its code to improve performance. However, after a second simulation run yielded even worse results with zero goals scored and an average of 7.3 goals conceded, the model saved its notes and submitted the policy.

MiniMax M3 analyzed the soccer engine's physics, match loop, and player formations before writing an initial 725-line policy that struggled heavily against a high-pressing opponent due to excessive fouls. To address this, the model revised the policy to reduce aggressive tackling and improve ball retention, which successfully lowered fouls from 13.8 to 2.8 per match and improved performance against possession-based teams. Ultimately, despite remaining vulnerable to the "gegenpress" strategy, the model saved its notes and submitted this iterated policy.