Emergent Communication

Emergent communication, also called emergent language, studies communication protocols that arise between AI agents solving a shared task without a fixed vocabulary given in advance. The field began with referential games in multi-agent reinforcement learning: a sender and a receiver network had to coordinate on symbols to identify a target image, with no reward for matching human words (Lazaridou et al., 2016; Foerster et al., 2016). ICLR has run an Emergent Communication Workshop since 2018, and surveys such as Lazaridou and Baroni (2020) and Peters et al. (2024) track the field.

Protocols that emerge this way tend to be compositional and morphologically productive: agents combine and inflect symbols systematically. They usually drift from human language, though, and can become opaque to observers. A 2026 study, GlossoGen (Stengel-Eskin et al., arXiv:2609.01491), extended this work to populations of LLM agents working under communication and time pressure. The invented shorthand these agents produced was passed on to newly added, weaker models, which learned it from usage alone. The authors call this generational transmission cumulative cultural evolution, a capacity for accumulating and refining shared conventions previously documented only in humans.

The finding matters beyond linguistics. As multi-agent LLM systems get deployed for coordination-heavy work such as trading or operations, protocols that drift toward efficiency rather than human readability limit how well people can monitor what agents are telling each other. Researchers treat this as a concrete, testable version of the broader oversight and interpretability problem, separate from claims about agent intent.