Emergent Communication in Multi Agent Referential Games

44 concepts43 papers1 proposal

How can independent artificial agents learn to coordinate and share meaning without a pre-existing language? The answer lies in a classic game-theoretic framework: the referential communication game.

In its simplest form, the game involves two agents: a speaker (or sender) and a listener (or receiver). The speaker is shown a specific target object—such as an image, a graph, or a collection of attributes—and must transmit a message consisting of discrete symbols chosen from a vocabulary. The listener receives this message and must correctly identify the target object from a lineup of distracting alternatives. Because the agents share a common reward based purely on the listener's success, they must co-evolve a shared protocol—an emergent language—to solve the task.

The Power of the Discriminative Bottleneck

To understand why referential games successfully generate meaningful communication, we must look at their core constraint: the discriminative objective.

In standard generative modeling or autoencoding, an agent is typically tasked with reconstructing an entire input signal perfectly. As Ueda notes, it remains an open question how pure reconstruction objectives translate to discriminative settings. When agents are forced to distinguish a target from distractors, they cannot simply copy low-level pixel noise or transmit an uncompressed bitstream. Instead, the pressure to differentiate forces the speaker to isolate features that uniquely identify the target relative to its context.

This mechanism acts as a powerful catalyst for semantic grounding. As Naszádi et al. explain, this discriminative objective forces agents to develop a shared vocabulary directly tied to the underlying features of the environment. Remarkably, this communicative pressure is so effective at extracting structure that it can double as an unsupervised representation learning technique. For instance, Dessi et al. demonstrated that when visual encoders are trained entirely from scratch via a referential game, they "accidentally" learn high-quality visual features that rival dedicated self-supervised contrastive learning frameworks like SimCLR.

How Perception Dictates the Code

While agents can easily achieve perfect task success in these games, the nature of the language they develop depends heavily on how they perceive their environment. Left unconstrained, neural networks are notorious opportunists. Bouchacourt et al. revealed that agents' internal visual representations can drift away from human-like conceptual hierarchies, allowing them to successfully communicate and solve games using inscrutable, low-level patterns—even when the inputs are random noise.

To steer agents toward human-like concepts, researchers manipulate the structure of the input space. The foundational intuition is that structured perception breeds structured language. Lazaridou et al. demonstrated that forcing agents to perceive the world as a collection of distinct, meaningful properties is a prerequisite for human-like, structured language to emerge.

Going a step further, Akkerman showed that changing the underlying representation from raw pixels to structured graphs (which explicitly encode spatial relations as edges) significantly enhances the compositionality of the emergent language. When the perceptual system is forced to bind features to distinct entities, the communication protocol naturally mirrors that compositional structure.

The Evolutionary Pressures for Structure

Human language relies on compositionality—the ability to combine reusable components (like "red" and "triangle") to describe entirely novel combinations. Why do agents choose compositionality over a holistic code where every unique object gets its own arbitrary name?

In a static, two-agent setup, compositionality is not strictly guaranteed. Chaabouni et al. found that deep learning agents can generalize to new concepts using highly entangled, non-compositional codes if the task performance alone is the only metric optimized. The pressure to build structured language instead comes from social and evolutionary constraints:

  1. Transmission and Imitation Bottlenecks: Languages must be easy to pass down. Cheng et al. discovered that when new generations of agents learn by imitating their predecessors, reinforcement learning acts as a natural filter that selects for structured, compositional languages because they are simply easier to learn and transmit. This aligns with the findings of Spike et al., who identified information loss and an anti-ambiguity bias as minimal requirements for establishing shared conventions.
  2. Environmental and Channel Noise: Real-world communication channels are imperfect. Verhoef et al. demonstrated that introducing transmission noise and incremental processing constraints forces neural listeners to develop human-like structural biases, such as minimizing dependency lengths. Under these efficiency pressures, language universals—like the trade-off between word order and case marking—emerge naturally (Lian et al.).

Crucially, researchers must be careful how they measure this structure. Conklin proved that emergent languages often exhibit natural-language-like variations, such as synonymy and flexible word orders, which traditional, rigid compositionality metrics mistakenly penalize as a lack of structure.

Scaling to Pragmatics and Game-Theoretic Equilibrium

In basic referential games, communication is purely literal. But human language operates on pragmatics—the ability to read between the lines and infer intent based on context.

To endow modern AI models with this capability, researchers frame communication as a strategic game where agents must actively model the listener's perspective. For example, Wu et al. utilized a referential game framework combined with preference optimization to teach vision-language models how to navigate context-aware communication and develop a rudimentary Theory of Mind.

Rather than relying on computationally expensive, infinite recursive reasoning ("I think that you think that I think..."), recent breakthroughs treat pragmatic communication as a search for a game-theoretic equilibrium. Jacob et al. introduced a framework (RECO) that computes a regularized equilibrium, balancing task success with a penalty for deviating from standard linguistic meanings.

This equilibrium-search perspective can even be applied to large language models (LLMs) without any additional training. By framing text generation as a consensus game between a generative agent and a discriminative agent, Jacob et al. demonstrated that searching for a strategic equilibrium between generation and scoring significantly improves an LLM's reasoning and accuracy.

Key Foundational Work

How agents see things: On visual representations in an emergent language game (2018) — contributed 1 concept elements to this direction.
Emergence of Linguistic Communication from Referential Games with Symbolic and Pixel Input (2018) — contributed 1 concept elements to this direction.
Multi-Agent Cooperation and the Emergence of (Natural) Language (2017) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_referential_communication_game_dd03ddce,
  title = {referential communication game},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/referential-communication-game/rdr_dd03ddce84cd45349b15e102b118daa1},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_referential_communication_game_dd03ddce_v1,
  title = {referential communication game},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/referential-communication-game/rdr_dd03ddce84cd45349b15e102b118daa1/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake