Predicting cost in a multi‑agent world model pushes most constrained policy updates into the imagination phase. Empirical results show that a stronger model‑free baseline requires $10$ to $20$ times the interactions of CA‑MAMBA on the median SMAC configuration to match its performance, while under equal budget CA‑MAMBA attains $2.8$ to $12$ times the MAMuJoCo reward of the baselines. In controlled tests that use a single task per benchmark, pricing the model’s cost advantage at parity cuts the cost of a cost‑blind learner by more than half and satisfies a tight continuous‑control limit that the learned multiplier fails to meet, suggesting that the model’s cost estimates are not the bottleneck.
Review