Metropolis–Hastings Naming Game (MHNG)

• MHNG is a decentralized communication framework that uses iterative Metropolis–Hastings updates to achieve emergent shared symbol systems.
• It applies a decentralized MCMC sampler with detailed balance to align agents' internal categorizations and sign assignments.
• Empirical studies show that MHNG improves clustering and convergence in both synthetic populations and human-machine interaction settings.

The Metropolis–Hastings Naming Game (MHNG) is a decentralized communication protocol and computational framework for symbol emergence between agents through iterative, role-alternating naming interactions. Each interaction stochastically proposes and accepts sign assignments for objects via a Metropolis–Hastings (MH) update step, targeting the central goal of distributed Bayesian inference over shared, latent symbolic variables. MHNG rigorously models emergent communication in both synthetic agent populations and human-machine dyads, enabling joint perceptual alignment and the construction of external symbol systems without explicit reward signals or centralized likelihood computation. It provides a principled methodology for studying and engineering co-creative learning, decentralized language evolution, and symbiotic AI alignment.

1. Formal Architecture and Mathematical Foundations

In MHNG, a pair of agents (typically labeled Speaker and Listener) coordinate over a set of $N$ discrete objects, each of which must be assigned a sign from a finite vocabulary. Each agent $m$ possesses: internal category assignments $c^{n,m}\in\{1,\ldots,K\}$, a sign–category mapping $\theta^m$ parameterizing $P(\text{sign}\mid \text{category})$, and optionally concept–observation parameters $\phi^m$. The protocol for an individual naming iteration is as follows (Okumura et al., 18 Jun 2025, Okumura et al., 2023, Taniguchi et al., 2022):

1. Proposal (Speaker): For object $n$, the Speaker samples candidate sign $s_n^*$ from their lexicon conditioned on their current category:

$$s_n^* \sim P(s_n \mid \theta^{Sp}, c_n^{Sp})\,, \quad \text{with}\; P(s_n^*=\ell) = \theta^{Sp}_{\ell, c_n^{Sp}}.$$

2. Acceptance (Listener): The Listener computes the Metropolis–Hastings acceptance ratio:

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

With probability $m$0, the Listener replaces their current sign $m$1 with $m$2; otherwise $m$3 is retained.

3. Updates: Both agents optionally update their parameters ($m$4, $m$5, $m$6) via local Gibbs sampling conditioned on their observation history.

The process alternates Speaker/Listener roles. The overall transition kernel satisfies a detailed-balance condition targeting the posterior over shared sign–category assignment latent variables given both agents' observations and hyperparameters. In this sense, the communication implements a decentralized MH sampler for the coupled generative model (Okumura et al., 2023, Taniguchi et al., 2022).

2. Probabilistic Generative Models and Joint Attention

MHNG presupposes a joint generative model (e.g., Interpersonal Gaussian Mixture, Inter-GM; Inter-MDM; Inter-GMM+VAE) governing the emergence of categories, signs, and percepts (Okumura et al., 18 Jun 2025, Furukawa et al., 2022). A general instantiation assumes:

• For each object $m$7, a shared sign $m$8 is drawn from a Categorical prior.
• Each agent samples an internal category $m$9 conditioned on $c^{n,m}\in\{1,\ldots,K\}$0 via their own $c^{n,m}\in\{1,\ldots,K\}$1.
• Each agent's observation $c^{n,m}\in\{1,\ldots,K\}$2 is generated from their category latent $c^{n,m}\in\{1,\ldots,K\}$3 via a modality-specific likelihood (e.g., in Inter-GM, $c^{n,m}\in\{1,\ldots,K\}$4 for category $c^{n,m}\in\{1,\ldots,K\}$5).

The full joint posterior targeted by the naming game is:

$c^{n,m}\in\{1,\ldots,K\}$6

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

MHNG communication occurs under joint attention: both agents attend to the same object during each step, enabling shared inference over the sign variable without revealing perception or model parameters.

3. Symbol Emergence as Decentralized Bayesian Inference

The central theoretical insight is that MHNG implements a decentralized MCMC sampler for the posterior $c^{n,m}\in\{1,\ldots,K\}$7, achieving emergent agreement over shared signs (i.e., "symbol emergence") as an emergent property of distributed Bayesian inference (Okumura et al., 2023, Taniguchi et al., 2022, Furukawa et al., 2022). The exactness and convergence of this protocol rest on standard MH-MCMC conditions (ergodicity, detailed balance, correct proposal and acceptance steps) and three communication requirements:

1. The Speaker samples from their local posterior over signs.
2. The Listener computes the acceptance probability via the local likelihood ratio for their current category.
3. Each agent performs Bayesian parameter updates (e.g., via Gibbs sampling) given its observation and the current sign assignments.

When these are satisfied, MHNG interaction non-increasingly reduces the expected KL divergence between the agents' joint beliefs and the true posterior, formalizing a notion of collective free-energy minimization (Okumura et al., 18 Jun 2025, Taniguchi et al., 2022). Each naming exchange incrementally aligns the agents' inference over latent symbolic variables, supporting a predictive-coding hypothesis for language emergence.

4. Extensions: Population Protocols and Model Variants

MHNG was originally formulated for dyadic (two-agent) systems, but recent work generalizes the method to multi-agent populations via the Recursive Metropolis–Hastings Naming Game (RMHNG) (Inukai et al., 2023). In the RMHNG, $c^{n,m}\in\{1,\ldots,K\}$8 agents recursively interleave MH updates: for each object, a sequence of MH steps incrementally integrates local evidence from each agent without centralizing the data. This recursive protocol provably yields an approximate Gibbs sampler for the full multi-agent joint posterior under standard regularity conditions.

Approximations—such as One-Sample (OS) and Limited-Length (LL) variants—significantly reduce computational cost while preserving fidelity to the Bayesian target. Empirical results demonstrate substantial gains in clustering and sign-sharing accuracy versus baseline protocols in both synthetic and real-image domains.

Structural variants of the generative model have also been explored:

• Head-to-head (H2H) Inter-MDM: The latent sign variable is modeled as a "child" of agents' internal categories, rather than a "parent," providing improved flexibility and facilitating extension to more complex sign structures and multimodalities without loss of convergence guarantees (Furukawa et al., 2022).
• Deep generative extensions: Integrating VAEs with GMMs enables application to high-dimensional sensory data, confirming that MHNG-based communication improves both emergent symbol sharing and unsupervised representation learning (Taniguchi et al., 2022).

5. Empirical Verification and Human Studies

Experimental work validates MHNG both in artificial and human-agent settings. In human-only studies, participants played the Joint Attention Naming Game (JA-NG) under MHNG protocols. Acceptance decision traces were compared across competing models (constant, numerator-only, heuristic subtractions) (Okumura et al., 2023). The MH-based acceptance probability significantly outperformed all baselines in predicting human acceptance/rejection actions, both in aggregate and per-individual analyses. Direct affine mappings of empirical acceptance to $c^{n,m}\in\{1,\ldots,K\}$9 revealed monotonic, significant correspondence.

In human-AI dyads, experiments confirmed that MHNG-enabled agents achieved statistically higher ARI (Adjusted Rand Index) for category assignments and greater convergence toward a shared sign system compared to always-accept or always-reject baselines (Okumura et al., 18 Jun 2025). Furthermore, the stochastic acceptance probability was mirrored in human listener responses, demonstrating humans' readiness to align with MHNG principles when collaborating with an AI partner.

Empirical evaluation has also highlighted that communication via MHNG-driven protocols leads to performance in emergent sign-sharing and categorization that approaches centralized Bayesian inference with only minimal information sharing.

6. Co-Creative Learning and Symbiotic AI Alignment

MHNG instantiates a new paradigm of co-creative learning, in which neither agent simply learns from the other, but both update internal models through principled, role-alternating interaction (Okumura et al., 18 Jun 2025). Unlike unilateral supervised learning or unsupervised clustering, the protocol supports emergent alignment across modalities or perception channels that are inherently partial or disjoint.

A key implication is that interfaces between humans and AI can leverage MHNG-style protocols for real-time, interactive alignment of perceptual models, facilitating concept or value convergence without requiring access to raw observations or gradients. This suggests a viable path toward the construction of genuinely symbiotic AI systems, in which humans and artificial agents develop integrated symbolic representations through decentralized, self-organizing communication. Future work is motivated toward extending these frameworks to richer concept spaces, population-scale language emergence, and continuous symbol systems.

7. Summary of Empirical Properties

The following table summarizes empirical results from human-human, human-AI, and multi-agent MHNG studies:

Setting	Communication Protocol	Clustering Accuracy (ARI)	Sign System Agreement ($\theta^m$0)	Notable Outcomes
Human-AI Dyad	MHNG	AI: 0.609, Human: 0.490	AI: 0.765, Human: 0.729	MHNG > always-accept/reject (Okumura et al., 18 Jun 2025)
Human-Human	MHNG vs. Heuristics	-	-	MHNG acceptance best predicts human behavior (Okumura et al., 2023)
MNIST Agents	MHNG + deep generative	0.78 (MHNG), 0.81 (Gibbs)	0.91 (MHNG), 0.0 (no comms)	Emergent symbol sharing, improved clustering (Taniguchi et al., 2022)
Multi-Agent	RMHNG	0.91 (synthetic), 0.61 (YCB)	0.99 (YCB)	RMHNG outperforms simplified/isolated baselines (Inukai et al., 2023)

Empirical evidence robustly supports the view that MHNG and its variants are effective mechanisms for decentralized Bayesian symbol emergence, with strong theoretical guarantees and demonstrable convergence in both artificial and human-involved systems.