In multi-agent reinforcement learning, agents that cooperate with unseen partners at test time benefit from training against a wide range of partner behaviors. A new paper on arXiv introduces PopGen, a method for generating such partner populations. The authors argue that existing approaches typically measure diversity using trajectory-level proxies, which may not capture the meaningful strategic differences that matter for coordination.

PopGen is designed to produce populations that are semantically diverse, meaning partners differ in the underlying goals or strategies they pursue, rather than only in the specific actions they take. According to the abstract, this shift in how diversity is generated should make trained cooperators more robust when paired with novel partners at test time.

The paper is announced as new on arXiv, but the abstract provided is partial, so details of the algorithm and experimental results are not yet available in this summary. The core claim, however, is clear: trajectory-level diversity is insufficient, and semantic diversity offers a better path to zero-shot coordination.