Scaling Emergent Communication to Large Agent Populations

52 concepts49 papers1 proposal

In the study of emergent communication, researchers look at how artificial agents can invent their own languages from scratch to solve collaborative tasks. For years, the standard experimental setup has been the dyad: two isolated agents playing a signaling game.

While dyadic setups are clean and easy to train, they suffer from a fundamental limitation. If two agents are stranded alone on an island, they can easily co-optimize a highly idiosyncratic, brittle shorthand that works perfectly for them but fails to generalize anywhere else. As Bouchacourt et al. point out in Miss Tools and Mr Fruit: Emergent Communication in Agents Learning about Object Affordances, training agents in isolated pairs fails to create the social pressure necessary to develop a stable, standardized language.

To build communication protocols that are robust, structured, and human-like, we have to scale the population. Moving from pairs to large communities introduces unique mathematical regularizers, structural challenges, and evolutionary dynamics.

Community Size as a Mathematical Regularizer

Why does a larger crowd force a better language? In a dyad, an agent can simply memorize its partner's specific quirks. But when an agent must communicate with hundreds of different peers, memorization fails. The agent is forced to generalize.

In Community size effect in artificial learning systems., Tieleman et al. demonstrated that larger agent communities act as a powerful regularizer. When multiple encoders and decoders are randomly paired, the resulting communication protocol shifts away from idiosyncratic codes toward highly abstract, generalizable representations.

This population-level pressure fundamentally alters how agents map their internal world to symbols. As Hoang et al. observe in SimSiam Naming Game: A Unified Approach for Representation Learning and Emergent Communication, scaling communication protocols past simple pairs forces a deep alignment of shared internal representations across the entire multi-agent system, preventing the models from drifting into private, unaligned semantic spaces.

The Threat of Dialect Fragmentation

While large populations incentivize abstraction, they also introduce a massive coordination problem: dialect divergence. In a large, decentralized network, agents cannot talk to everyone at once; they primarily interact with localized subgroups.

Early foundational work by Steels et al. in Situated Grounded Word Semantics warned that scaling to thousands of participants risks a failure to reach global communicative coherence, causing the population to fracture into regional dialects.

However, this fragmentation isn't necessarily a failure of coordination. Kim et al. investigated this exact trade-off in Emergent Communication under Varying Sizes and Connectivities. They discovered that as the number of agents scales up, message similarity across the global population does indeed decrease. Yet, surprisingly, the overall success of information exchange remains high. Large-scale multi-agent systems maintain functional coordination not by enforcing a single universal language, but by allowing diverse, overlapping local protocols to handle the heavy lifting.

[Agent A1] <---> [Agent A2] (Dialect A)
       \          /
        [Agent B1] <---> [Agent B2] (Dialect B)

Figure 1: In large networks, global uniformity drops, but overlapping local protocols allow information to flow successfully across the entire population.

When structural rules are carefully designed, global consensus can still emerge from the bottom up. Sierra-Santibáñez demonstrated in An Agent-Based Model of the Emergence and Transmission of a Language System for the Expression of Logical Combinations that even when noise and regional variations threaten to disrupt alignment, localized "language games" allow stable grammatical norms to gradually crystallize across hundreds of interacting agents over multiple generations.

Algorithmic Catalysts for Decentralized Consensus

How do we mathematically guarantee that a large population can reach a consensus without a central coordinator? One elegant solution is to frame decentralized communication as a distributed inference problem.

In Recursive Metropolis-Hastings naming game: symbol emergence in a multi-agent system based on probabilistic generative models, Inukai et al. proved that scaling the classic two-agent naming game to a population of $N$ agents is mathematically equivalent to performing approximate Bayesian inference over a shared latent variable. By using a recursive probabilistic acceptance mechanism, agents can incrementally align their vocabularies globally while only ever interacting in pairs.

But algorithm design is only half the battle; the environment must also demand complexity. Simply broadcasting messages to a larger audience does not automatically make a language more structured. Lee et al. proved this in One-to-Many Communication and Compositionality in Emergent Communication, showing that structured, compositional language only emerges at scale when listeners have diverse interests or are forced to coordinate their actions collectively.

When vocabulary sizes scale to thousands of terms, unconstrained natural language can suffer from severe linguistic drift. To stabilize this process, Zhang introduced SIGN: Schema Induced Games for Naming (Student Abstract), demonstrating that enforcing a minimal structural schema on messages allows massive agent populations to reach consensus up to 5.8× more effectively while drastically reducing token overhead.

Horizontal Pressure and Efficient Compression

Finally, scaling communication requires a shift in how we think about language transmission. Many classical models of language evolution focus heavily on vertical transmission—passing language down from a single teacher to a single student over generations.

Vertical Transmission:   [Teacher] ---> [Learner] (Prone to information loss)

Horizontal Pressure:     [Peer A] <===> [Peer B]  (Forces efficient compression)

In Culturally transmitted color categories in LLMs reflect a learning bias toward efficient compression, Imel et al. argue that relying solely on vertical transmission neglects the crucial role of horizontal communicative pressure—the immediate, peer-to-peer social interaction between contemporary agents.

When hundreds of peers must actively understand and be understood by one another in real time, horizontal pressure forces the language to compress efficiently. The vocabulary naturally restructures itself to balance simplicity (cognitive ease for the speaker) with precision (clarity for the listener), allowing highly scalable and robust semantic systems to emerge across the entire community.

Key Foundational Work

Sophisticated Cooperation in FA/C Distributed Problem Solving Systems (1991) — contributed 1 concept elements to this direction.
Using Joint Responsibility to Coordinate Collaborative Problem Solving in Dynamic Environments (1992) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 49 papers
  2. Version 1 · 2026-10-04 · 80 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_multi_agent_communication_scaling_16c8059f,
  title = {multi-agent communication scaling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-agent-communication-scaling/rdr_16c8059f22d14038bd68623cb9763687},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_multi_agent_communication_scaling_16c8059f_v2,
  title = {multi-agent communication scaling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-agent-communication-scaling/rdr_16c8059f22d14038bd68623cb9763687/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake