Tag
Microsoft's NextLat introduces a training objective that rewards belief-state representations instead of relying solely on next-token prediction, pushing models toward compact world models for better generalization.