Metropolis-Hastings Naming Game (MHNG)

• MHNG is a computational framework where agents use decentralized Bayesian inference and Metropolis-Hastings proposals to align their symbol systems.
• It leverages local observations with probabilistic generative models and iterative exchanges to robustly converge on shared symbols.
• Empirical results on datasets like MNIST and Fruits360 and human–AI experiments demonstrate significant improvements in clustering accuracy and inter-agent agreement.

The Metropolis-Hastings Naming Game (MHNG) is a computational framework for emergent communication and symbol alignment, grounded in the mathematics of decentralized Bayesian inference and Metropolis-Hastings (MH) Monte Carlo. Initially formulated for two agents and extended to multi-agent populations, MHNG rigorously models how agents—artificial or human—interactively construct and align discrete symbol systems by exchanging candidate signs and updating beliefs based only on local, partial information. The MHNG provides not only an algorithmic basis for the emergence of shared lexica but also a framework for analyzing empirical semiotic behavior and co-creative learning in human, artificial, and hybrid dyads (Taniguchi et al., 2022, Okumura et al., 18 Jun 2025, Okumura et al., 2023, Inukai et al., 2023, Furukawa et al., 2022).

1. Conceptual Foundations: Decentralized Bayesian Inference through Naming Games

MHNG formalizes emergent communication as an instance of approximate decentralized Bayesian inference. Agents interact through a repeated naming game where, under joint attention, they refer to the same object and communicate a candidate sign without receiving any explicit reward or feedback. The core innovation is that symbol emergence (the process by which agents converge on shared signs for objects or categories) is not treated as supervised or reinforcement learning, but rather as probabilistic inference over latent variables in a joint generative model conditioned on all agents’ observations (Taniguchi et al., 2022).

Each agent possesses a perceptual representation derived from its local sensory data and maintains parameters for generative and prior distributions over percepts and symbols. Communication proceeds by proposing and probabilistically accepting or rejecting candidate signs according to the Metropolis-Hastings rule, such that the interactive process as a whole implements a distributed Markov chain Monte Carlo sampler for the joint posterior of all latent variables (Taniguchi et al., 2022, Okumura et al., 18 Jun 2025, Okumura et al., 2023).

2. Mathematical Formulation and Algorithmic Structure

Probabilistic Generative Model (PGM). For an object indexed by $d$:

• $o^A_d, o^B_d$: Observations of agents $A$, $B$
• $z^A_d, z^B_d$: Latent perceptual variables
• $w_d$: Shared discrete sign

The joint generative model for two agents factorizes as: $$p(o, z, w, \theta, \phi) = p(\theta^A)p(\theta^B)p(\phi^A)p(\phi^B) \prod_{d=1}^D p(w_d) p(z^A_d | w_d, \phi^A) p(o^A_d | z^A_d, \theta^A) p(z^B_d | w_d, \phi^B) p(o^B_d | z^B_d, \theta^B)$$ where $p(w_d)$ is a categorical prior.

Metropolis-Hastings Protocol: Let the speaker propose $w^\star$ sampled from its local posterior. The listener accepts this proposal with probability

$$r = \min\Big(1, \frac{P(z^{Li}_d | w^\star, \phi^{Li})}{P(z^{Li}_d | w_d^{old}, \phi^{Li})}\Big)$$

or, in categorical models,

$$r_n^{MH} = \min \left(1, \frac{P(c_n^{Li} | \theta^{Li}, s^\star)}{P(c_n^{Li} | \theta^{Li}, s^{Li})}\right)$$

Both agents then update their parameters ($\theta$, $\phi$) and latent variables via local Gibbs or variational steps, conditioned on their data and the accepted sign (Taniguchi et al., 2022, Okumura et al., 2023, Okumura et al., 18 Jun 2025, Furukawa et al., 2022).

Pseudocode Example:

procedure MH_COMMUNICATION(z_sp, φ_sp, z_li, φ_li, w_li) w_sp ∼ P(w | z_sp, φ_sp) r = min(1, P(z_li | w_sp, φ_li) / P(z_li | w_li, φ_li) ) if Uniform(0,1) ≤ r: return w_sp else: return w_li end

Repeated alternation of these exchanges implements a Monte Carlo process sampling from the joint posterior over signs given both agents’ data.

3. Variants and Model Instantiations

a. Inter-GMM+VAE Model

MHNG is instantiated in deep generative architectures such as the inter-Gaussian Mixture Model plus Variational Autoencoder (inter-GMM+VAE). Each agent models its observation as

$$w_d \sim \mathrm{Cat}(\pi), \quad z_d^* | w_d = k \sim \mathcal{N}(\mu_k^*, \Lambda_k^{*-1}), \quad o_d^* | z_d^* \sim \text{Decoder}_{\theta^*}(z_d^*)$$

where agent-specific GMM parameters serve as priors for the VAE latent code. The inference protocol alternates MH-based label exchanges with Gibbs and variational parameter updates, resulting in mutual inference over the shared signs (Taniguchi et al., 2022).

b. Interpersonal Multimodal Dirichlet (Inter-MDM) Models

MHNG was formulated for both tail-to-tail and head-to-head types of Inter-MDM for multi-agent, multimodal data:

• Tail-to-tail: The sign variable is a global latent shared directly between agents.
• Head-to-head (H2H): The sign variable is sampled as a product-of-experts function over agent-internal category assignments.

Empirical results demonstrate nearly identical and robust emergence of shared symbol systems under both configurations (Furukawa et al., 2022).

c. Multi-Agent Generalization: Recursive Metropolis-Hastings Naming Game (RMHNG)

MHNG extends to N agents as RMHNG, recursively chaining MH exchanges: the proposal for agent $n+1$ is drawn from the posterior generated by the n-agent chain. Efficient approximations (one-sample, limited-length) yield major computational savings with minimal empirical loss of accuracy (Inukai et al., 2023).

4. Convergence, Theoretical Guarantees, and Bayesian Properties

MHNG inherits the convergence properties of Metropolis-Hastings: the process satisfies detailed balance with respect to the true joint posterior, and—under suitable support guarantees and appropriate proposal distributions—the chain is irreducible and aperiodic, ensuring convergence to the unique stationary distribution (Taniguchi et al., 2022, Okumura et al., 2023, Okumura et al., 18 Jun 2025, Furukawa et al., 2022, Inukai et al., 2023). In multi-agent and partial observability settings, the listener always conditions only on local data and its own current beliefs, ensuring decentralized operation. The collective free energy (KL divergence to the true posterior plus evidence constant) is non-increasing in expectation, establishing Monte Carlo consistency (Okumura et al., 18 Jun 2025).

Table: Convergence Properties of MHNG

Property	Theoretical Guarantee	Reference
Detailed Balance (2-agent)	Yes	(Taniguchi et al., 2022, Furukawa et al., 2022)
Ergodicity/Irreducibility	Yes (Dirichlet or GMM prior)	(Furukawa et al., 2022, Okumura et al., 18 Jun 2025)
Bayesian Consistency	KL divergence non-increasing	(Okumura et al., 18 Jun 2025)
Stationary Distribution	Joint sign/categorization	(Okumura et al., 2023, Inukai et al., 2023)

5. Empirical Validation: Symbol Alignment and Cross-Modal Inference

MHNG has been empirically validated using synthetic, real-image, and human–AI/AI–AI experiments:

• On MNIST (upright/rotated) and Fruits360, the inter-GMM+VAE MHNG protocol increased agent ARI (clustering accuracy) from ≈0.65 (no communication) to ≈0.78, with inter-agent label agreement κ≈0.91—nearly matching the centralized Gibbs topline ARI≈0.81 (Taniguchi et al., 2022).
• In YCB object view-differentiated settings, RMHNG achieved ARI≈0.61, κ≈0.99, and posterior matching rates ≈76% relative to centralized sampling (Inukai et al., 2023).
• Co-creative human–AI naming games (MHNG) achieved the highest ARI for AI (0.609±0.246), and sign histogram agreement (≈0.77 for AI, ≈0.73 for humans), showing integration of partial observations and robust symbol convergence (Okumura et al., 18 Jun 2025).
• Human listeners’ acceptance behavior is closely predicted by the precise MH acceptance probability, outperforming constant, heuristic, and binary models (pooled-data U-tests: all p < 10^{-3}) (Okumura et al., 2023).

6. Extensions, Applications, and Cognitive Implications

The recursive MHNG (RMHNG) generalizes the protocol to N-agent decentralized symbol emergence, with tunable complexity for large populations (Inukai et al., 2023). Both exact and one-sample/limited-length approximations preserve strong performance in clustering and symbol alignment. The formal equivalence of MHNG-based emergent communication to decentralized (collective) predictive coding and cross-modal inference establishes its relevance to studies of language evolution, multimodal AI alignment, and empirical semiotics (Taniguchi et al., 2022, Okumura et al., 18 Jun 2025, Okumura et al., 2023).

A central empirical finding is that both human–AI and human–human dyads align their sign-acceptance behavior quantitatively with the theoretical MH rule, evidencing decentralized Bayesian inference as a plausible model of human symbolic coordination (Okumura et al., 18 Jun 2025, Okumura et al., 2023).

7. Summary and Outlook

MHNG provides a principled, empirically validated algorithmic account of emergent communication—enabling the distributed formation, sharing, and recall of symbolic systems among artificial, human, and hybrid multi-agent collectives. The formal properties of decentralized Monte Carlo inference guarantee robust convergence and alignment, with demonstrated applicability to diverse modalities and heightened relevance for co-creative, symbiotic AI systems (Taniguchi et al., 2022, Okumura et al., 18 Jun 2025, Okumura et al., 2023, Furukawa et al., 2022, Inukai et al., 2023). The model’s grounding in both mathematical theory and experimental semiotics positions it as a foundational protocol in computational studies of symbol emergence.