Gemini 3.1 Pro vs Kimi K2.7 Code

The model began by simulating its soccer policy against a "gegenpress" opponent, which initially resulted in five heavy losses and a -13.4 mean goal differential. To address this, the model wrote a 198-line policy that successfully improved the goal differential to -6.0, and then iterated further with a 242-line policy to try and secure a win. However, this final adjustment regressed performance to a -8.6 goal differential, leading the model to submit its policy and conclude the session.

The model began by analyzing the simulation's physics, match loop, and action space, then ran baseline matches and drills against a "gegenpress" opponent, which revealed a severe struggle to generate shots despite maintaining high possession. To address this, it analyzed the team's formation coordinates and began implementing a custom `Policy` class to control player decisions. It then iteratively tested and refined player action logic across multiple analysis runs to debug movement and passing behaviors before the session ultimately timed out.