An Oxford University study posted to arXiv on November 4, 2026, finds that large language model agents coordinate poorly with partners they have never worked with before. Even in minimal test environments, including the Cat & Dog signaling game, frontier agents underperformed because of limited understanding of coordination and poor reasoning about their partners’ beliefs. The researchers found that shared rules written in plain natural language improve performance by encouraging safe fallback actions and consistent conventions. These coordination-friendly definitions can also be discovered automatically through meta-learning, matching or outperforming hand-crafted rules. Whether the approach holds up in complex real-world settings remains an open question.
Source: Read the original article
