A new arXiv paper introduces CA-MAMBA, a constrained multi-agent reinforcement learning approach that leans on a multi-agent world model. The key idea is to predict cost inside that world model, which lets the algorithm carry out most constrained policy updates in imagination rather than through direct environment experience.
The authors report that on the median StarCraft Multi-Agent Challenge (SMAC) configuration, a stronger model-free baseline requires 10 to 20 times more interactions than CA-MAMBA. That is a substantial efficiency gap, suggesting that imagination-based cost prediction can dramatically cut the sample burden for constrained multi-agent tasks.
Because only the abstract is available, the claim is limited to interaction counts on median SMAC settings; the paper does not yet provide full performance comparisons or broader benchmarks in this summary. Still, the result points to a promising direction for making constrained multi-agent RL more practical.