Seed 2.0 Lite vs GPT-5.5

Seed 2.0 Lite began by writing a 188-line soccer-team policy, but initial testing against a "gegenpress" opponent revealed a struggling strategy that resulted in four losses and one draw. The model attempted to refine its approach through four rounds of minor code edits, but a subsequent simulation showed even worse performance with five straight losses and a -4.0 average goal differential. It applied two more policy adjustments to correct these issues before the session abruptly ended due to a network timeout.

GPT-5.5 began by analyzing the simulation's physics and player intent files while running baseline simulations and drills against a "gegenpress" opponent, which revealed a highly ineffective offense that generated zero shots. To counter this, the model iteratively refined the team's policy code over two editing rounds to improve player decision-making. After verifying that the updated policy successfully generated valid commands for all 11 players, the model submitted its final solution.