Metropolis-Hastings Naming Game: Decentralized Inference

• The Metropolis-Hastings Naming Game is a computational model for emergent communication that uses a decentralized MH protocol to align symbols among agents.
• It applies probabilistic generative modeling and Bayesian inference to transform inter-agent exchanges into a distributed MCMC sampling problem without explicit rewards.
• Empirical studies demonstrate high categorization accuracy and robust convergence in human, AI, and multiagent scenarios, highlighting its practical impact on symbol emergence.

The Metropolis-Hastings Naming Game (MHNG) is a formal computational model of emergent communication among agents, grounded in probabilistic generative modeling and decentralized Bayesian inference. It frames the process of symbol emergence—the development and alignment of shared symbolic representations—as a decentralized Markov chain Monte Carlo (MCMC) sampling problem, where agents use a Metropolis–Hastings (MH) protocol to coordinate on shared signs (symbolic labels) while encoding their own partial, private observations of objects. The MHNG paradigm formally captures both human–human and human–AI communication as the iterative proposal, acceptance, and updating of candidate signs for objects under joint attention, without explicit reward, supervision, or feedback beyond acceptance decisions. This mechanism provably aligns agents’ internal representations, yielding convergent symbol systems and improved unsupervised learning, and generalizes naturally to multi-agent and mixed human–AI scenarios (Taniguchi et al., 2022, Okumura et al., 2023, Okumura et al., 18 Jun 2025, Inukai et al., 2023).

1. Probabilistic Generative Modeling and Symbol Emergence

At the core of the MHNG is the formulation of agent communication via a joint probabilistic generative model (PGM). Consider $N$ agents, each observing $D$ shared objects. For object $d$, agent $n$ observes $x_d^n$, with the assumption that all observations are conditionally independent given a discrete latent variable $w_d$ (the "sign" or symbol assignment):

$$w_d \sim \operatorname{Cat}(\gamma), \qquad x_d^n \mid w_d, \theta^n \sim p(x \mid w_d, \theta^n), \quad \forall n = 1,\ldots,N.$$

More expressive forms, such as inter-agent Gaussian mixture models (Inter-GMMs) combined with deep latent-variable autoencoders (Inter-GMM+VAE), extend this setup, enabling rich category and feature formation. Each agent learns its own generative parameters ($\theta^n, \phi^n$) independently; only the signs $\{w_d\}$ couple the PGMs (Taniguchi et al., 2022, Inukai et al., 2023).

Symbol emergence is recast as Bayesian inference of shared latent variables (signs), given only inter-agent exchanges of signs, not raw data. For $N$ agents, the objective is to approximate the joint posterior

$$p(\{w_d\}_{d=1}^D, \{\theta^n\}_{n=1}^N \mid \{x_d^n\}).$$

This structure enables decentralized, feedback-free learning of symbol systems, with no need for explicit reward or back-channel signals.

2. Metropolis–Hastings Protocol for Decentralized Inference

The defining feature of the MHNG is the implementation of inter-agent communication as a distributed Metropolis–Hastings algorithm. Each naming interaction consists of:

1. Speaker Proposal: The Speaker samples a candidate sign $w'$ for object $d$ from its local posterior $p(w|x_d^{\mathrm{Sp}}, \theta^{\mathrm{Sp}})$.
2. Listener Acceptance: The Listener evaluates the acceptance ratio

$$\alpha = \min\left\{1, \frac{p(x_d^{\mathrm{Li}} | w', \theta^{\mathrm{Li}})}{p(x_d^{\mathrm{Li}} | w, \theta^{\mathrm{Li}})}\right\},$$

and either updates $w_d^{\mathrm{Li}} \leftarrow w'$ (accept) or retains the previous sign (reject).

1. Parameter Update: Each agent updates its own generative model parameters using local Gibbs sampling or variational methods, never accessing other agents’ observations or parameters.

This cyclical process is repeated with alternating Speaker and Listener roles across all objects and agents, forming a decentralized blocked Gibbs sampler with MH substeps for the shared latent variables (Taniguchi et al., 2022, Okumura et al., 18 Jun 2025, Inukai et al., 2023).

3. MHNG as Decentralized Bayesian Inference and Collective Predictive Coding

Rigorous theoretical analysis shows that the MHNG implements a distributed MCMC sampler for the joint posterior over signs:

$$p(\{w_d\} | \{x_d^n\}, \{\theta^n\}) \propto \prod_{d=1}^D p(w_d) \prod_{n=1}^N p(x_d^n | w_d, \theta^n).$$

Each MH interaction maintains detailed balance with respect to the target posterior, guaranteeing convergence of the agents’ shared lexicon to samples from the true joint posterior under mild regularity conditions (irreducibility, aperiodicity).

A key conceptual insight is the collective predictive coding hypothesis: communication via the MHNG constitutes society-wide free-energy minimization across distributed brains/agents, with each agent's VAE or GMM trained to minimize variational free energy (the evidence lower bound, ELBO), and the MH sign exchange effectively extending predictive coding across individuals. This aligns symbol emergence, sign negotiation, and multimodal representation learning under a unified, principled Bayesian framework (Taniguchi et al., 2022, Okumura et al., 18 Jun 2025).

4. Algorithmic Realizations and Multi-agent Extensions

The basic two-agent MHNG can be realized as follows:

for iter in range(I): for d in range(D): # Speaker->Listener for dth object w_star = sample_posterior_w(x_d_speaker, theta_speaker) accept_prob = min(1, p(x_d_listener | w_star, theta_listener) / p(x_d_listener | w_current, theta_listener)) if uniform(0,1) <= accept_prob: w_d_listener = w_star # Parameter update: each agent updates theta^n theta_speaker = gibbs_sample_theta(x_speaker, w_speaker) theta_listener = gibbs_sample_theta(x_listener, w_listener) # Roles swap and repeat

The generalization to $N$ agents, termed the Recursive MHNG (RMHNG), applies a recursive, tree-structured or chain MH proposal and acceptance sequence. The proposal for $w_d$ is constructed recursively using MH communication among a subset of agents, and the final acceptance is based on the listener’s likelihood. The RMHNG provably converges to the full Bayesian posterior, leveraging decentralized chains of proposal and acceptance steps. Two computationally efficient approximations are introduced: the one-sample (OS, $T=1$) and limited-length (LL) reductions, both empirically shown to accelerate convergence while yielding symbol emergence (Inukai et al., 2023).

5. Empirical Studies: Human, AI, and Mixed Dyads

Extensive empirical work demonstrates the viability and generality of the MHNG framework:

• Human–Human Experiments: In laboratory Joint Attention Naming Game (JA-NG) settings, human dyads synchronize naming policies over repeated rounds, making accept/reject decisions per the MH acceptance probability. Empirical regression confirms listeners' decisions are significantly aligned with MH-computed probabilities, rejecting constant-probability null models. The MH-based model predicts human acceptance more accurately than heuristic rivals (Okumura et al., 2023).
• Human–AI Co-creative Learning: Online experiments with human–MHNG-agent pairs under partial observability show that MH-based negotiation yields higher categorization accuracy and stronger sign convergence than always-accept or always-reject baselines. Binned human acceptance rates closely track theoretical MH probabilities, and ARI as well as sign agreement with the full-data posterior are maximized in MHNG-aligned interactions (Okumura et al., 18 Jun 2025).
• Multi-agent Symbol Emergence: Synthetic and real-image data experiments confirm that RMHNG approaches centralized Gibbs in ARI and sign-sharing metrics, both for Gaussian mixtures and high-dimensional object images, while OS and LL approximations balance computational efficiency with communication efficacy (Inukai et al., 2023).
• Deep Generative Architectures: The Inter-GMM+VAE MHNG model achieves high ARI and sign-sharing on image data (MNIST, Fruits360), forming robust, interpretable categories, and enabling reconstruction from uttered signs, outperforming no-comm and all-accept baselines (Taniguchi et al., 2022).

Study Domain	Agent Type(s)	Key Metrics	MHNG Results
Human–Human (JA-NG)	Human–Human	Acceptance, precision	MH-based model outperforms Constant, Numerator, Subtraction, Binary (Okumura et al., 2023)
Human–AI Dyads	Human–AI	ARI (cat. accuracy), sign agreement	MH computer ARI ≈0.61, sign agreement ≈0.77, both higher than AA/AR (Okumura et al., 18 Jun 2025)
Multi-agent	Synthetic, YCB images	ARI, κ, sign alignment	RMHNG ARI ≈0.90, sign alignment ≈96% to Gibbs (Inukai et al., 2023)
Deep Emergence	MNIST, Fruits360	ARI, κ, reconstruction quality	ARI ≈0.78, κ ≈0.92, clear prototypes (Taniguchi et al., 2022)

6. Theoretical Significance and Connections to Cognitive Science

The MHNG operationalizes communication as decentralized Bayesian inference, viewing the negotiation and acceptance of signs as cross-modal, interpersonal inference. It eschews explicit rewards, feedback, or centralized supervision in favor of minimalistic, feedback-free joint attention. The process realizes “collective predictive coding” and “free-energy minimization” at the societal level, providing a computational foundation for the emergence of symbol systems in both artificial and natural agents.

Human experiments confirm that natural dyads organize symbol systems in a manner consistent with MHNG, suggesting the model captures cognitive substrates of lexical alignment. The potential for mixed-agent populations supports research avenues in human–robot and human–AI communication, with extensions to more ecological modalities and open-ended settings anticipated (Okumura et al., 2023, Okumura et al., 18 Jun 2025).

7. Extensions, Limitations, and Future Directions

Recent developments include the RMHNG for $N$-agent systems, scalable approximations (OS, LL), and deep generative model instantiations. Future research targets include incorporating Bayesian nonparametric models for open sign spaces, compositional symbol systems, richer perceptual domains, and ecological communication settings. Integration with symbiotic AI frameworks—where humans and AI mutually co-create shared symbolic representations—signals emerging directions for alignment, social representation learning, and the computational study of language evolution (Inukai et al., 2023, Taniguchi et al., 2022, Okumura et al., 18 Jun 2025).

Empirical limitations pertain to the constrained categorical structure, forced-choice paradigms, and restricted stimuli in existing human and hybrid experiments. Further studies with naturalistic stimuli, unconstrained communication protocols, and larger multi-agent or human–machine collectives will elucidate the universality and limitations of the MHNG framework.