Tag
This paper proposes a minimal 'advantage-style' action channel for latent world models that cancels action-independent distractor variation by subtracting the mean effect over actions, improving controllability without auxiliary losses or reconstruction.