Metropolis–Hastings Naming Game

• Metropolis–Hastings Naming Game is a decentralized probabilistic framework where agents use local Bayesian inference to converge on shared symbolic names.
• It embeds Metropolis–Hastings sampling within interactive naming protocols, enabling agents to update and refine symbol assignments based solely on local information.
• Empirical results in human and human–AI experiments show that MHNG improves categorization accuracy and aligns with predictive coding principles in multi-agent communication.

The Metropolis–Hastings Naming Game (MHNG) is a decentralized, probabilistic framework for symbol emergence in populations of interacting agents. It formalizes how agents—whether human, artificial, or a mix—achieve consensus on naming or symbolic systems under joint attention, by embedding Metropolis–Hastings (MH) steps within an interactive communication protocol. MHNG posits that agents do not simply memorize shared signs but approximate joint Bayesian inference over latent symbolic variables, leveraging only local information and exchange of signs. This paradigm has unified explanatory and constructive power, linking cognitive semiotics, multi-agent systems, and emergent communication with decentralized Bayesian modeling and Markov chain Monte Carlo methodologies.

1. Probabilistic Generative Model for Symbol Emergence

MHNG is underpinned by a probabilistic generative model (PGM) that describes how agents perceive, categorize, and propose names for objects. For each object $n=1,\ldots,N$, a shared latent symbolic variable (sign) $s_n$ is postulated, which generates each agent’s perceptual category $c_n^m$ conditioned on agent-specific parameters $\theta^m$, and ultimately the agent’s raw observation $x_n^m$ parameterized by $\phi^m$:

$$\begin{aligned} s_n &\sim p(s \mid \gamma) \ c_n^m &\sim p(c \mid s_n, \theta^m), \quad m\in\{1,\ldots,M\} \ x_n^m &\sim p(x \mid c_n^m, \phi^m) \end{aligned}$$

Here, $p(s|\gamma)$ usually represents a uniform prior over a discrete set of labels; $\theta^m$ specifies mapping strength between signs and internal concepts; $\phi^m$ encodes perceptual emission (e.g., parameters of a Gaussian mixture model) (Okumura et al., 2023, Okumura et al., 18 Jun 2025, Taniguchi et al., 2022, Inukai et al., 2023).

2. Mechanism: The Metropolis–Hastings Step in Naming Games

Each round of the naming game involves role allocation into speaker (Sp) and listener (Li), with the following protocol:

1. Speaker’s proposal: Sp proposes a sign $s_n^*$ for object $n$, sampled from their predictive posterior $$Q(s_n^*\mid c_n^{\mathrm{Sp}}) = P(s_n^* \mid \theta^{\mathrm{Sp}}, c_n^{\mathrm{Sp}})$$.
2. Listener’s acceptance: Li, upon receiving $s_n^*$, computes the MH acceptance ratio:

$$r_n^{\mathrm{MH}} = \min\left(1, \frac{P(c_n^{\mathrm{Li}} \mid \theta^{\mathrm{Li}}, s_n^*)}{P(c_n^{\mathrm{Li}} \mid \theta^{\mathrm{Li}}, s_n^{\mathrm{Li}})}\right)$$

Li samples a uniform random variable $u \sim \mathrm{Uniform}(0,1)$. If $u \le r_n^{\mathrm{MH}}$, they accept the proposal and update their assignment $s_n^{\mathrm{Li}} \leftarrow s_n^*$; otherwise, they retain their own sign (Okumura et al., 2023, Okumura et al., 18 Jun 2025, Taniguchi et al., 2022). Both agents then update their internal parameters via local Bayesian updates or Gibbs sampling.

This procedure implements a decentralized Metropolis–Hastings MCMC sampler targeting the global posterior $P(\{s_n\} \mid \{x_n^m\}_{m=1}^M)$, ensuring detailed balance across the interacting population (Inukai et al., 2023).

3. Decentralized Bayesian Inference and Predictive Coding

MHNG realizes decentralized Bayesian inference: agents, while exchanging only symbolic assignments, collectively approximate the global posterior over shared symbolic variables. This protocol operates without direct transmission of raw data or learning gradients—crucial for privacy, heterogeneity, or cross-modal settings.

The theory generalizes to $N$ agents in the Recursive Metropolis–Hastings Naming Game (RMHNG), wherein proposal chains are recursively passed and accepted across the agent population with each applying local MH ratios based on their private data and parameters. The acceptance step for each new agent in the chain is:

$$\alpha = \min \left\lbrace 1, \frac{p(x^{n+1} | w', \theta^{n+1})}{p(x^{n+1} | w, \theta^{n+1})} \right\rbrace$$

This yields an ergodic chain with stationary distribution matching the joint posterior (Inukai et al., 2023).

The collective effect is the emergence of a shared symbolic system (naming scheme) that aligns with the joint inference over the data and internal models of all agents. This operationalizes predictive coding at the population level and supports the "collective predictive coding hypothesis" for symbol emergence and language evolution (Taniguchi et al., 2022).

4. Human and Human–AI Experiments

Empirical validation of MHNG includes semiotic studies with human dyads (Okumura et al., 2023) and human–AI pairs (Okumura et al., 18 Jun 2025). For example:

• Human dyads: Participants played JA-NG rounds in which one proposed a name for a color patch and the other chose to accept or reject. Human acceptance rates of proposals were well-explained by a biased Bernoulli model with acceptance probabilities linearly related to MH-derived $r_n^{\mathrm{MH}}$ (fit: $a \approx 0.51$, $b \approx 0.48$). The MHNG acceptance model outperformed constant or heuristic (e.g., numerator-only, subtraction, binary) baselines in predicting participant behavior.
• Human–AI dyads: Experiments under partial observability (human and AI obtained different features of the same object) revealed that h-AI pairs with an MH-based agent achieved higher categorization accuracy (AI-side ARI: $0.609 \pm 0.246$ for MH vs $0.469 \pm 0.207$ for always-accept) and stronger convergence to the correct sign posterior than control agent types. Moreover, human acceptance decisions closely tracked $r_n^{\mathrm{MH}}$ (sensitivity $\hat a = 0.645 \pm 0.300$), substantiating the cognitive plausibility of MHNG.

A summary of experimental contrasts is given below.

Variant	Human Behavior Fit	Category Accuracy (AI-side)	Sign Agreement
MH-based	High	$0.609 \pm 0.246$	$0.765 \pm 0.069$
Always Accept	Lower	$0.469 \pm 0.207$	$0.717 \pm 0.073$
Always Reject	Lower	$0.404 \pm 0.223$	$0.469 \pm 0.054$

(Okumura et al., 18 Jun 2025, Okumura et al., 2023)

5. Algorithmic Implementations and Deep Generative Instantiations

MHNG has been instantiated with both shallow (Gaussian mixture) and deep generative models (e.g., VAE+GMM for real images). In deep setups, each agent employs a VAE to encode raw observations $o_d^*$ into latent variables $z_d^*$, which are linked to the sign variable $w_d$ via a GMM prior. The MH naming game coordinates assignment of $w_d$ without direct sharing of latent or data representations (Taniguchi et al., 2022).

Experimental results on MNIST and Fruits360 indicate that mutual-inference MHNG approaches match or approach centralized Bayesian inference in clustering quality (ARI) and sign agreement (Cohen's $\kappa$), outperforming no-communication or all-acceptance baselines.

On synthetic and real-image multi-agent tasks, extensions such as RMHNG and its computationally scalable variants (One-Sample, Limited-Length) enable symbol emergence across N agents, with posterior-matching rates exceeding 96% for full RMHNG and substantial efficiency gains with approximations (Inukai et al., 2023).

6. Theoretical Properties and Broader Significance

MHNG, and its recursive extensions, possess formal guarantees: under mild assumptions, they converge to the (possibly marginalized) global joint posterior over sign variables and agent parameters, thus implementing decentralized Gibbs or Metropolis–Hastings samplers across the agent population.

This decentralized inference framework provides a rigorous formalization for semiotic communication, symbol emergence, and language evolution, without requiring reward feedback, global coordination, or transmission of high-dimensional percepts. MHNG bridges cognitive, algorithmic, and engineering perspectives by:

• Explaining experimental semiotic phenomena in humans as decentralized Bayesian inference
• Enabling constructive recipes for emergent communication in human–AI and multi-agent robotic systems
• Connecting symbol emergence to predictive coding and collective free-energy minimization principles

A plausible implication is that by encoding the MHNG protocol into artificial agents or robots, true co-creative learning and robust symbiotic AI alignment can be realized under distributed, partial observability conditions (Okumura et al., 18 Jun 2025).

7. Extensions and Ongoing Research

Recent work has generalized MHNG to N agents (RMHNG), introduced scalable approximations (One-Sample, Limited-Length chain), and empirically demonstrated robust, high-accuracy symbol emergence in both synthetic and real-image domains with feature extraction via pre-trained deep networks (Inukai et al., 2023).

MHNG, as a model and as an algorithmic paradigm, continues to inform research in emergent communication, decentralized learning, multi-agent systems, experimental semiotics, and AI alignment. The integration of deep probabilistic modeling and interactive naming game protocols is an active area, especially for cross-modal grounding, continual adaptation, and symbol negotiation at population scale.

References:

• (Okumura et al., 2023)
• (Okumura et al., 18 Jun 2025)
• (Taniguchi et al., 2022)
• (Inukai et al., 2023)