[Paper Review] Emergent Multi-Agent Communication in the Deep Learning Era

Angeliki Lazaridou, Marco Baroni|arXiv (Cornell University)|Jun 3, 2020

Language and cultural evolution107 references32 citations

TL;DR

This paper surveys how deep-learning agents develop emergent communication in cooperative and competitive settings, analyzes how to study and interpret these languages, and discusses avenues to improve AI coordination and human–machine interaction.

ABSTRACT

The ability to cooperate through language is a defining feature of humans. As the perceptual, motory and planning capabilities of deep artificial networks increase, researchers are studying whether they also can develop a shared language to interact. From a scientific perspective, understanding the conditions under which language evolves in communities of deep agents and its emergent features can shed light on human language evolution. From an applied perspective, endowing deep networks with the ability to solve problems interactively by communicating with each other and with us should make them more flexible and useful in everyday life. This article surveys representative recent language emergence studies from both of these two angles.

Motivation & Objective

• Motivate study of language emergence as a way to understand human language evolution and to create flexible, interactive AI.
• Review the progression from simple referential games to complex, perceptually rich environments with deep agents.
• Discuss analytical methods for decoding emergent protocols and assessing genuine communication.
• Explore how emergent communication can improve inter-agent coordination and enable better human–machine interaction.

Proposed method

• Describe continuous versus discrete communication and their implications for learning and back-propagation.
• Summarize representative studies using referential games, multi-turn interactions, and embodied 3D environments.
• Review analytical metrics for signaling, listening, and causal influence to establish genuine communication.
• Discuss measures of compositionality and their relation to generalization and disentanglement in representations.
• Outline approaches to align emergent languages with natural language through supervised grounding, iterated learning, and pre-trained models.

Experimental results

Research questions

• RQ1 Under what conditions do deep-agent communities develop communicative protocols with real informational content?
• RQ2 How can we quantify and verify genuine communication versus degenerate strategies in emergent languages?
• RQ3 What factors promote compositionality and generalization in emergent protocols?
• RQ4 How can emergent communication be leveraged to improve inter-agent coordination and human–machine collaboration?

Key findings

• Emergent languages often rely on discrete bottlenecks and may diverge from human-like semantics, requiring careful analysis to confirm meaningful communication.
• Continuous communication generally enhances coordination, while discrete channels can struggle in complex environments without specialized training.
• Compositionality in emergent languages is not guaranteed and may emerge under specific biases, representations, or community dynamics.
• Human–machine interaction benefits from grounding emergent languages in natural language or supervision to reduce language drift and improve interpretability.
• Inter-agent coordination and pragmatic biases can be strengthened by incorporating social or cooperative incentives and grounded communication objectives.
• When humans interact with machines using natural-language-based cheap talk, coordination improves compared to non-language communication.