Lombardi

We sought to build the best NFL coach the world has ever seen. We built an adversarial RL environment where two LLMs compete as NFL coordinators across full simulated drives. Every play is resolved by a two-stage world model trained on 16,000+ real NFL plays from the 2024 season from the NFL Big Data Bowl dataset. In our world model an outcome classifier predicts what happens (normal play, touchdown, interception, fumble), then quantile regressors sample realistic yardage conditioned on that outcome. Two Qwen2.5-1.5B models were LLMs Interactions are captured via sklearn's HistGradientBoostingClassifier and HistGradientBoostingRegressor. Coverage schemes affect pass outcomes, blitz rates influence sack probability, and formation matchups produce realistic yard distributions, among others. The action space is hierarchical and constrained.