Drawing with Strangers: Population Scaling Drives Zero-Shot Mutual Intelligibility in Emergent Sketching

Researchers demonstrate that scaling training populations in emergent communication systems enables zero-shot mutual intelligibility (ZMI)—successful communication between independently trained agent groups with no prior exposure. The study uses emergent sketching as a communication modality, showing that larger populations develop universal visual-grounding strategies rather than closed-group dialects, with potential applications for building interoperable AI systems.

This research addresses a fundamental challenge in multi-agent AI systems: creating communication protocols that work across independently developed populations. Traditional emergent communication research focuses on generalization within a single group or to novel inputs, but this work explores a more demanding scenario—genuine zero-shot communication between strangers from completely separate training cohorts.

The key insight involves population scaling as a mechanism for achieving universality. Rather than agents developing idiosyncratic shortcuts optimized for their specific group, larger populations force agents to converge on objective visual features of target images. This perceptual grounding acts as a natural coordination point, allowing cross-group communication without explicit agreement. The mechanism mirrors how humans from different cultures can recognize and describe sketches based on shared visual reality.

For AI development, this finding carries significant implications for building socially interoperable systems. Current AI deployment often assumes homogeneous training and deployment contexts, but real-world applications require interaction across diverse AI systems developed independently. The result suggests that scaling and population diversity during training may inadvertently create more transferable communication protocols.

The research also reveals that increased in-group variation accompanies population scaling, indicating that universality doesn't require homogeneity or loss of expressiveness. Future work should explore whether these principles apply to other emergent communication modalities beyond sketching and whether they scale to human-AI interaction scenarios.

• →Scaling training populations in multi-agent systems drives convergence toward universal communication strategies based on objective visual features.
• →Zero-shot mutual intelligibility between independently trained groups increases substantially with population size, without requiring prior inter-group exposure.
• →Larger populations maintain higher in-group communicative variation while reducing cross-group variation, achieving universality through perceptual grounding rather than homogenization.
• →The findings suggest a pathway for developing AI agents capable of interoperability across independently developed systems.
• →Population diversity during training inadvertently creates more transferable and generalizable communication protocols.