Symbol Emergence through Metropolis Hastings Naming Games

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 220 papers31 conceptsfrozen 2026-10-04

How do independent agents—whether humans or robots—agree on a shared language without a central authority or a pre-existing dictionary? This process, known as symbol emergence, requires agents to map their private, internal sensory categories to a public set of arbitrary signs.

Recent research led by Tadahiro Taniguchi and Yoshinobu Hagiwara suggests that this social coordination is not merely a heuristic process of imitation, but a mathematically rigorous form of decentralized Bayesian inference. By framing communication as a sampling process, these researchers have shown how a stable lexicon can emerge from the local interactions of agents who never have direct access to each other's "mental" states.

Communication as Distributed Inference

In traditional AI, communication is often viewed as a data transfer problem. In the symbol emergence framework, however, communication is viewed as a way for multiple agents to perform joint inference over a shared environment.

Imagine two robots looking at the same object. Robot A sees a "red sphere" and Robot B sees a "round toy." They have different sensors and internal categorizations. To agree on a name, they must find a symbol that maximizes the likelihood of both their observations. Yoshinobu Hagiwara and colleagues demonstrate in Multiagent Multimodal Categorization for Symbol Emergence that this interaction can be modeled as a Metropolis-Hastings (MH) algorithm.

In this "Naming Game," the exchange of signs follows a specific probabilistic protocol:

  1. The Speaker (proposer) generates a sign based on its internal posterior distribution of categories.
  2. The Listener (acceptor) evaluates this sign against its own internal model.
  3. The Acceptance Rule: The listener accepts or rejects the sign based on an acceptance ratio.

This mechanism ensures that the agents effectively perform Markov Chain Monte Carlo (MCMC) sampling. Over many iterations, the signs they exchange converge to represent a shared latent variable—a "symbol"—that bridges their individual experiences.

The Intuition of Probabilistic Acceptance

Why use a complex acceptance-rejection rule instead of simple imitation? Ryota Okumura and Tadahiro Taniguchi found that human pairs playing naming games follow this Bayesian logic more closely than simple reinforcement or imitation Metropolis-Hastings algorithm in joint-attention naming game.

The probabilistic nature of the MH Naming Game allows the system to avoid "local optima"—where a group might get stuck with a poor or ambiguous name—and ensures that the emerging lexicon is robust to the "noise" of different sensory perspectives. As shown in Emergent Communication through Metropolis-Hastings Naming Game with Deep Generative Models, this approach allows agents with deep generative models (like VAEs) to synchronize their internal latent spaces through the bottleneck of symbolic signs.

Scaling to Populations

While the basic MH Naming Game works for pairs, human language involves thousands of agents. Jun Inukai and Akira Taniguchi extended this to larger groups in the Recursive Metropolis-Hastings Naming Game (RMHNG).

In this recursive structure, the consensus reached by a small sub-group acts as the "proposal distribution" for the next agent. This allows the entire population to converge on a shared symbol system without a central controller. Mathematically, the researchers proved that this decentralized sign exchange is equivalent to performing distributed Bayesian inference on a shared latent variable across the entire network.

Collective Predictive Coding and System 3

The implications of this work extend beyond simple naming. Tadahiro Taniguchi proposes the Collective Predictive Coding (CPC) hypothesis, which views language as a "System 3"—a collective cognitive layer that exists above individual brains Beyond individuals: Collective predictive coding for memory, attention, and the emergence of language.

In this view:

By treating communication as a decentralized inference process, this research provides a unified framework for understanding how robots can learn from each other, how humans align their concepts, and how collective intelligence emerges from local, probabilistic interactions.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 15 papers
  2. Version 1 · 2026-10-04 · 20 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_decentralized_bayesian_inference_0ec9c04d,
  title = {decentralized Bayesian inference},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/decentralized-bayesian-inference/rdr_0ec9c04d9fe24d24a48018b6b753fcd5},
  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_symbol_emergence_through_metropolis_hast_0ec9c04d_v1,
  title = {Symbol Emergence through Metropolis Hastings Naming Games},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/symbol-emergence-through-metropolis-hastings-naming-games/rdr_0ec9c04d9fe24d24a48018b6b753fcd5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}