Generative Emergent Communication: Large Language Model is a Collective World Model (2501.00226v1)

Abstract: This study proposes a unifying theoretical framework called generative emergent communication (generative EmCom) that bridges emergent communication, world models, and LLMs through the lens of collective predictive coding (CPC). The proposed framework formalizes the emergence of language and symbol systems through decentralized Bayesian inference across multiple agents, extending beyond conventional discriminative model-based approaches to emergent communication. This study makes the following two key contributions: First, we propose generative EmCom as a novel framework for understanding emergent communication, demonstrating how communication emergence in multi-agent reinforcement learning (MARL) can be derived from control as inference while clarifying its relationship to conventional discriminative approaches. Second, we propose a mathematical formulation showing the interpretation of LLMs as collective world models that integrate multiple agents' experiences through CPC. The framework provides a unified theoretical foundation for understanding how shared symbol systems emerge through collective predictive coding processes, bridging individual cognitive development and societal language evolution. Through mathematical formulations and discussion on prior works, we demonstrate how this framework explains fundamental aspects of language emergence and offers practical insights for understanding LLMs and developing sophisticated AI systems for improving human-AI interaction and multi-agent systems.

• The paper presents a novel framework where emergent communication unifies world models and LLMs through collective predictive coding.
• It introduces a mathematical formulation that reinterprets language as a generative model, contrasting conventional mutual information approaches.
• The findings highlight practical implications in multi-agent reinforcement learning by demonstrating language emergence via decentralized Bayesian inference.

Generative Emergent Communication: LLM as a Collective World Model

This paper, authored by Tadahiro Taniguchi and colleagues, proposes a unifying theoretical framework called generative emergent communication (EmCom) which bridges emergent communication, world models, and LLMs through the concept of collective predictive coding (CPC). The framework posits that the emergence of symbol systems, particularly language, can be understood as processes of decentralized Bayesian inference across multiple agents via a generative approach. This framework extends beyond the conventional discriminative model-based approaches predominant in emergent communication studies.

Key Contributions and Theoretical Foundations

The paper makes significant contributions in two key areas. Firstly, it introduces generative EmCom as a new framework, demonstrating how the emergence of communication in multi-agent reinforcement learning (MARL) settings derives from the perspective of control as inference (CaI). This formulation parallels the notions of predictive coding and the free-energy principle, suggesting a theoretical alignment between individual cognitive processes and societal language evolution. Secondly, it provides a mathematical formulation interpreting LLMs as collective world models. This conceptualization rests on integrating multiple agents' experiences through CPC, suggesting that language formed via this process inherently encodes information about the world as observed collectively by human agents.

Generative EmCom Versus Conventional Approaches

The generative EmCom approach differs fundamentally from conventional signaling games used in emergent communication research, which rely mainly on maximizing mutual information between signal inputs and outputs. Instead, generative EmCom regards language as a generative model, incorporating these signals as latent variables in its structure. This framework aligns with the principles of active inference, where language naturally emerges to minimize free energy or prediction errors within a society, collectively maximizing environmental predictability.

The Metropolis-Hastings naming game (MHNG) is presented as an instance of this concept, highlighting decentralized Bayesian inference as the mechanism driving symbol emergence. Here, the MHNG operates as a decentralized sampling method whereby agents coordinate on a shared symbol system through joint attention and language usage, further validating the CPC hypothesis.

Implications for LLMs and World Models

In discussing practical applications, the paper emphasizes the relevance of the proposed framework to LLMs. It argues that LLMs, trained on linguistic data, can be understood as embodying a collective world model informed by human sensorimotor experiences. This interpretation not only elucidates why LLMs possess latent structural knowledge about the world but also implies that such models capture a shared symbolic representation of multimodal experiences, enabling a deeper understanding of the emergent language and its link to world modeling.

The extension to MARL contexts emphasizes how communication via generative EmCom facilitates cooperative behavior among agents, leveraging shared latent structures to enhance predictive capabilities. Integrating world models within this framework allows agents to effectively align their internal states with actions that reinforce cooperative multi-agent strategies, thus emphasizing the practical utility of this conceptual unification.

Future Directions

The proposed paradigm provides a robust theoretical basis for exploring emergent communication, language evolution, and world modeling in AI. Future research might focus on empirical validation of the collective world model hypothesis, the dynamics of language evolution across generations, and the further integration of multimodal sensory-motor data in expanding the capabilities of LLMs. By leveraging generative principles, this framework holds promise for advancing our understanding of AI systems capable of sophisticated human-like communication and environmental adaptation.