New research reveals LLMs can coordinate like humans without talking

A new study published on arXiv (v2 June 2026) by researchers from Oxford University and Bar-Ilan University explores how large language models (LLMs) coordinate in multi-agent settings without explicit communication—a critical capability for real-world AI deployments like human-AI collaboration and safety systems. The paper, titled 'Tacit Coordination of Large Language Models,' evaluates over 20 LLMs (including models from major providers and open-source alternatives) across cooperative and competitive scenarios, including simulated search-and-rescue operations.

The team found that LLMs demonstrate a remarkable ability to coordinate through 'focal points'—salient solutions that emerge naturally without direct communication—often matching or even outperforming human coordination in these tasks. However, the same models consistently underperformed in scenarios requiring numerical reasoning or culturally specific contextual understanding, revealing social limitations in current LLM architectures. The researchers also tested simple, learning-free strategies that significantly improved coordination not only among LLMs but also between humans and AI systems. These findings underscore a critical gap: LLMs may not inherently share humans' cultural or perceptual frameworks, which could lead to misalignment in real-world applications where tacit coordination is essential.