PROWL

PROWL is a novel reinforcement learning-driven adversarial framework where an RL agent explores game environments with the objective of exposing failures in a world model and generating new training data from those failures.

The core idea is that world models accumulate blind spots when trained on limited data — situations the model handles poorly because they were rare or absent in training. An adversarial RL agent is trained specifically to discover these failure modes by reward-seeking in regions where the world model is least accurate.

Discovered failures are then used as training data to improve the world model, closing the gap between model capabilities and environmental complexity. This creates a virtuous cycle where world model improvement enables more capable adversarial agents, which in turn find more failure modes.

PROWL represents a shift from passive data collection toward active, targeted generation of training signal — addressing the coverage limitations of recorded demonstrations in world model training.