Kimi K2.7 Code vs Qwen 3.7 Max

Kimi K2.7 Code analyzed the simulation's physics, match loop, and player action space before running a baseline test against a "gegenpress" opponent, which resulted in five heavy losses and a -7.8 average goal differential. The model then developed and debugged a custom 451-line, 11-player policy to improve performance. Testing the new policy against the same opponent revealed a slight defensive improvement, reducing the average goal deficit to -6.2, after which the model saved its notes and submitted the code.

After analyzing the physics engine to address a weak baseline, Qwen 3.7 Max wrote a 637-line policy implementing a "stop before shooting" strategy to maximize kick accuracy. However, testing revealed this approach backfired against high-pressing opponents, dropping possession and increasing conceded goals because players held the ball too long. The model iterated by adjusting this logic, ultimately submitting a policy that, while still struggling against the "gegenpress" style, successfully improved its record against the "possession" opponent to 4W-1D-0L.