GPT-5.5 vs MiMo v2.5 Pro

The model began by analyzing the simulator's physics, intent resolver, and match loop, testing its initial policy against a "gegenpress" opponent where it struggled with low possession and a losing record. It then iteratively edited the team policy over several rounds, using drills and match simulations to evaluate its changes. While these adjustments successfully improved buildout possession from 34% to 58% in drills, match results against the gegenpress remained tough at 1W-0D-5L, leading the model to save its final policy and submit.

After analyzing the game's physics and player command structures, MiMo v2.5 Pro ran baseline simulations and discovered its policy was losing all five matches against a "gegenpress" opponent while conceding 5.2 goals per game. In response, the model implemented a 442-line policy overhaul focused on defensive compactness and direct attacking. Testing the new policy showed improved defensive resilience by cutting conceded goals to 2.6 and securing a draw (0-1-4 record) before the session timed out.