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

A preprint posted to arXiv on 9 June 2026 reports that training artificial agents in larger populations markedly improves their ability to communicate with strangers from entirely separate communities, a capability the authors term zero-shot mutual intelligibility [1][2]. The study, titled "Drawing with Strangers," examines emergent sketching, a visually grounded communication method where agents convey concepts through sequences of drawn strokes [1][2]. The researchers formalize zero-shot mutual intelligibility (ZMI) as successful communication between independently trained populations that have had no prior exposure to one another [1][2]. A central finding is that scaling the training population size increases in-group communicative variation, which prevents the agents from co-adapting into a single, homogeneous protocol [1][2]. At the same time, cross-group variation decreases as population size grows, indicating a structural convergence toward a form of universality [1][2]. The paper argues that this universality is achieved through perceptual grounding: larger populations anchor their emergent sketches more closely on the objective visual resemblance of the target images [1][2]. The work positions ZMI as a distinct axis of generalization in emergent communication, separate from prior research that focused on novel inputs or linguistic structures [1][2]. The authors suggest these findings outline a route toward building socially interoperable artificial agents [1][2]. The preprint was submitted on 9 June 2026 and is available on arXiv, with associated code and media linked through platforms including Hugging Face [1]. Emergent communication research has historically concentrated on how agents develop shared languages within a single group. The new paper shifts attention to inter-group understanding without any shared training history [1][2]. The results indicate that simply increasing the number of agents in the initial training pool yields communication protocols that are more recognizable to outsiders, without requiring explicit coordination between groups [1][2]. The analysis draws on the visual nature of sketching, where the physical resemblance of a drawing to its referent provides a grounding signal that scales with population diversity [1][2]. The authors report that this perceptual anchor grows stronger as the training population expands, driving the observed convergence in cross-group variation [1][2].