Gemini 3.5 Flash vs MiMo v2.5 Pro

The model began by analyzing the game's physics and player intent mechanics to address a severe -9.2 goal differential against a "gegenpress" opponent. It ran defensive and build-out drills to construct a custom 11-player policy, which initially narrowed the goal differential to -6.4 and improved shot production. However, a subsequent policy iteration regressed to a -7.8 goal differential with zero shots, and the session timed out while the model was running further diagnostic analyses.

MiMo v2.5 Pro analyzed the soccer simulation's physics, player commands, and match loop before running initial tests that revealed the existing policy was getting dominated by "gegenpress" opponents (0-5 record, -4.4 goal differential) due to excessive fouls. To counter this, the model wrote a 441-line policy designed to handle heavy pressing and reduce fouls. Testing the new policy showed improved resilience against the gegenpress (rising to 1W-1D-3L and -3.2 GD) while maintaining stable performance against possession, prompting the model to save its notes and submit the code.