Cooperative success often depends on implicit communication: partners converge on conventions, or regularities that give an action meaning beyond its direct effect. But conventions are rarely observed directly. In a new arXiv preprint, the authors propose a method for identifying and quantifying these hidden regularities in both humans and AI agents.
The paper frames conventions as something that can be located and measured, rather than merely inferred from outcomes. The abstract notes that conventions are usually observed only indirectly, for example through a score drop when they are broken. The new approach appears to formalize this intuition, offering a way to pinpoint what the convention is and where it operates in a cooperative interaction.
Because the work covers both human and AI agents, it could provide a common language for studying how different kinds of partners build shared understanding. This matters for designing AI systems that can coordinate with people in settings where explicit communication is limited or impossible.
This summary is based on a single source, so there are no conflicting reports to compare. The full paper is available on arXiv under the title "What and Where in Identifying and Quantifying Conventions in Humans and AI Agents."