Concept-Guided Spatial Regularization for World Models in Atari Pong

arXiv cs.AI · 6 hours ago

Researchers evaluated five visual world-model agents in Atari Pong by freezing their learned models and testing them with separate policies, finding failures like ball disappearance and incorrect motion across all models. In pixel-space zero-shot reinforcement learning, DreamerV3's performance dropped from -5.5 to -20.9 mean return when policies were trained entirely within frozen world models. The authors proposed Concept-Guided Spatial Regularization to improve world models by applying auxiliary reconstruction loss to task-critical regions, which improved results for three of five models but showed variable effects across the others.