TL;DR
This paper explores how language-like communication protocols emerge among cooperative agents in a multi-agent foraging environment, demonstrating properties similar to natural language through deep reinforcement learning.
Contribution
It introduces a framework for studying language emergence in multi-agent systems, highlighting how ecological and social factors influence language properties.
Findings
Agents develop language with properties like arbitrariness and compositionality
Emergent language is shaped by social dynamics and environmental constraints
The framework enables analysis of language evolution in embodied multi-agent settings
Abstract
Language is a powerful communicative and cognitive tool. It enables humans to express thoughts, share intentions, and reason about complex phenomena. Despite our fluency in using and understanding language, the question of how it arises and evolves over time remains unsolved. A leading hypothesis in linguistics and anthropology posits that language evolved to meet the ecological and social demands of early human cooperation. Language did not arise in isolation, but through shared survival goals. Inspired by this view, we investigate the emergence of language in multi-agent Foraging Games. These environments are designed to reflect the cognitive and ecological constraints believed to have influenced the evolution of communication. Agents operate in a shared grid world with only partial knowledge about other agents and the environment, and must coordinate to complete games like picking up…
Peer Reviews
Decision·Submitted to ICLR 2026
1. The introduction of the Foraging Games (FG) effectively integrates ecological constraints such as partial observability and the necessity of temporal reasoning. This moves beyond traditional, more static RefGame and offers a more ecologically valid setting for studying language origins. 2. The paper provides clear evidence for the emergence of both Interchangeability, arbitrariness, compositionality, cultrual transmission and, more notably, Displacement—the ability to refer to non-present f
Fundamental Critique on Research Significance and Paradigm: This is the most critical weakness. The research paradigm employed—utilizing small LSTM models in a highly simplified, "toy" $5 \times 5$ environment—appears to be meaningless. In the age of LLMs which have demonstrated scalable compositionality and robust generalization, the conclusions drawn from this limited setting face major challenges in terms of relevance and transferability. This work provides little inspiration for improving a
- The investigation of network structure on properties of the emergent communication protocol are novel and very interesting in my opinion. I think this is the main contribution of the paper, and would suggest reframing the paper around it considering that many of the other ideas are not novel (see Weaknesses). - The paper is well-situated in a growing body of literature on emergent communication. In particular, the game is similar to the setting of Mordatch and Abbeel (2017) which was a seminal
- The Foraging Game setup is very similar to the setup explored by Mordatch and Abbeel (2017), in what is a foundational work in emergent communication. This was the first (or among the first) papers to study embodied emergent communication, specifically with agents moving on a (continuous) two dimensional world, without seeing each other, in order to reach landmarks (equiv. pick up objects) at a certain order. That paper explicitly studied compositionality. None of this is mentioned in the pape
The paper is well motivated from the viewpoint of using this experimental setup for studying human language evolution. The paper is well situated within the related literature. The proposed framework of studying emergent communication in foraging games is compelling and conceptually intuitive. The results give hints that this framework would lead to desirable phenomena to be reflected (e.g., population size effects). The paper comes with a rich appendix, studying important factors such as vo
Contribution. Foraging games / 2d grid worlds have been extensively studied in multi-agent reinforcement learning (as noted in the submitted version). I am surprised that this would be (as claimed) new to the field of emergent communication (?) My rating is based on the assumption that foraging was not studied in emergent (symbolic) communication before. Baselines. The paper proposes a new experimental framework to study emergent communication. In this setting, a suitable baseline would be the
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
