Language Games in AI and Communication

Updated 4 March 2026

Language games are rule-governed communicative protocols defined by agents, message sets, state transitions, and reward functions.
They enable systematic exploration of emergent communication, compositionality, and pragmatic reasoning in both human and artificial agents.
Research in language games informs multi-agent dialogue, reconstruction tasks, and game-theoretic strategies to enhance coordination and learning.

A language game is a rule-governed communicative protocol in which agents—human or artificial—use language as a strategic instrument to achieve specific objectives. Stemming from Wittgenstein's philosophical investigations, language games encompass a diverse field of models and experiments, ranging from emergent communication in artificial agents to logical/mathematical inference, grounded interactive learning, and large-scale sociotechnical systems. This article synthesizes key formal frameworks, empirical findings, and theoretical debates in contemporary research on language games.

1. Foundational Definitions and Formalizations

The foundational conception of a language game, following Wittgenstein and later formalizations, is a protocol specifying (i) a set of players or agents, (ii) a message space (typically sequences over a finite alphabet), (iii) state and transition rules regulating conversation history, (iv) an interaction policy for each agent, and (v) terminal scoring functions that define the utility or reward for each player at the end of the game (Schaul, 2024, Guo et al., 2020, Aryan, 30 Jan 2025).

A generic formalization frames a language game as a tuple

$G = (P,\, \mathcal{M},\, \mathcal{S},\, T,\, R)$

where $P$ is the agent set, $\mathcal{M}$ the message set, $\mathcal{S}$ the state space, $T$ the transition function (mapping current state and joint messages to next state), and $R = (R^1, ..., R^n)$ the terminal scoring functions for each agent. The protocol—which may encode turn-taking, role assignments, private information, and termination conditions—provides the mechanics through which meaning, coordination, and strategy are enacted (Schaul, 2024, Aryan, 30 Jan 2025, Nevens et al., 2020).

In mathematical contexts, a language game is defined as a rule-governed practice in which moves are sentence-utterances constrained by syntactic and inferential rules, with proofs as “plays” comprising sequences of such moves (Landsman et al., 2023).

2. Taxonomy of Language Games: Classes and Objectives

Language games can be systematically categorized along several axes, reflecting both their functional objectives and methodological design (Guo et al., 2020, Suglia et al., 2023, Nevens et al., 2020, Schlangen, 2023):

Game Family	Communication Objective	Example Task(s)
Discriminative	Select correct output (classification, reference)	Referential games, VQA, visual entailment
Generative	Generate outputs (description, captioning)	Image captioning, dense captioning
Interactive	Goal-oriented dialogue, sequential planning	Dialogue, negotiation, command execution
Reconstruction	High-fidelity input reproduction	Signal reconstruction, emergent code transfer
Verdict/Conversation	Strategic discourse leading to adjudication	Turing test, court interrogation
Mathematical	Inferential, proof-based rule-following	Formal proofs, theorem checking

Discriminative and generative games typically emphasize one-shot or open-loop signal interpretation or generation, while interactive and verdict games instantiate multi-step, turn-based (possibly adversarial) dialogue with strategic reasoning and feedback. Mathematical language games formalize “moves” as inferential steps in accordance with logical calculus (Landsman et al., 2023).

In emergent communication, the referential game requires a sender to emit a message about an input $x$ that allows the receiver to disambiguate $x$ from distractors. The reconstruction game tasks the receiver with high-fidelity reconstruction of the sender’s input from a message (Guo et al., 2020).

3. Inductive Biases, Compositionality, and Expressivity

Recent empirical research demonstrates that the structure of the language game imposes characteristic inductive biases on emergent communication protocols (Guo et al., 2020). Specifically:

Referential games bias toward compositional languages, partitioning the input space along discriminative features and yielding high topographic similarity (Pearson correlation between input space and message space distances, converging to ≈0.45).
Reconstruction games yield higher information expressivity but lower compositionality (TS ≈0.15). Expressivity is quantified via performance transfer to downstream tasks, revealing that reconstruction-emerged protocols are more generally informative, while referentially-emerged protocols are more abstract and compositional.

A significant trade-off arises: maximizing compositionality may sacrifice full information content and vice versa. Statistical tests confirm robust differences in the emergent protocol properties ( $p$ -values for TS and transfer performance $<10^{-29}$ and $5.10 \times 10^{-64}$ , respectively) (Guo et al., 2020).

4. Learning, Grounding, and Pragmatic Reasoning

The emergence and learning of language within games rely on interaction-driven protocols, pragmatic inference, and adaptive policies (Wang et al., 2016, Nevens et al., 2020, Khani et al., 2018). In interactive environments such as the SHRDLURN blocks-world, agents learn mappings from utterances to actions/instructions through feedback on task success, without pre-specified lexica. Key findings:

Humans facilitate learning by using compositional, consistent lexica and avoiding synonyms, leading to faster system adaptation (Wang et al., 2016).
Pragmatic modeling via the Rational Speech Acts (RSA) framework enables agents to reason about the informativeness and intent underlying utterances, thus accelerating convergence and disambiguation in sequential games (Wang et al., 2016, Khani et al., 2018).
In grounded games, language emerges through the negotiation of form–meaning mappings over repeated referential tasks, with agents developing continuous perceptual ontologies and lexical alignments in physical or simulated environments (Nevens et al., 2020).

Evaluation of such games employs communicative success rates, ontological and lexical inventory metrics, and compositionality measures (Nevens et al., 2020, Guo et al., 2020).

5. Strategic Communication and Game-Theoretic Analysis

Game-theoretic models provide deep explanatory power for language games involving strategic interaction, private information, and iterative dialogue (Aryan, 30 Jan 2025, Khani et al., 2018, Xu et al., 23 May 2025). Conversation games are formalized as (possibly extensive-form) multi-agent games in which utterances constitute actions and payoffs depend on downstream judgments (e.g., a non-strategic judge in a verdict game). Solution concepts such as subgame-perfect and perfect Bayesian equilibria structure the analysis of optimal policies, deception, and information transfer.

Key empirical findings include:

Strategic agents utilizing look-ahead search and role-aware planning outperform naive agents in adversarial conversation games (e.g., prosecution vs. defense, Turing test imitation), with performance improvements exceeding 2× in win rate over naive baselines ( $p \le 10^{-5}$ ) (Aryan, 30 Jan 2025).
Metaphor-driven language games (e.g., CoMet framework) illustrate the use of covert, asymmetric communication channels, requiring dynamic metaphor generation and opponent modeling within multi-agent settings (Xu et al., 23 May 2025).

6. Language Games in Multi-Agent, Societal, and Open-Ended Settings

On a larger scale, language games underpin frameworks for open-ended exploration, continual self-improvement (Socratic learning), and the modeling of knowledge emergence in sociotechnical systems (Wen et al., 31 Jan 2025, Schaul, 2024). Critical mechanisms include:

Role fluidity: Dynamic reallocation of agent roles produces distributional diversity in data and trajectories.
Reward variety: Vector-valued, structured reward functions foster broadening of capability spectra.
Rule plasticity: Iterative adaptation or evolution of rules maintains system novelty and guards against stagnation (data reproduction trap).

These principles facilitate expanded data reproduction, in contrast to closed-loop retraining, and are adopted in frameworks aiming for artificial superhuman intelligence or fully recursive language-based self-improvement (Wen et al., 31 Jan 2025, Schaul, 2024). Agent capacity, coverage of games, and aligned feedback are necessary conditions for unbounded performance improvements. However, scaling brings profound challenges around alignment, feedback informativeness, time complexity, and avoidance of misalignment-driven drift (Schaul, 2024).

7. Language Games, Mathematical Structure, and Hybrid Models

Advanced theoretical work seeks to synthesize social-pragmatic (language game) and algebraic-geometric (semantic field theory, transformer-based) accounts of language (Vartziotis, 1 Jan 2026). While transformers and large-scale LLMs instantiate regularities explained by field-theoretic principles—continuous embeddings, attention as geometric kernels—critical aspects such as indexicality, pragmatic implication, social norm inference, and context sensitivity remain out of reach for purely mathematical models. Language games supply the social and pragmatic grounding—the boundary conditions—that set the parameters for geometric structure, indicating a fundamental complementarity between social and mathematical perspectives on language (Vartziotis, 1 Jan 2026).

References

(Guo et al., 2020) Inductive Bias and Language Expressivity in Emergent Communication.
(Wang et al., 2016) Learning Language Games through Interaction.
(Aryan, 30 Jan 2025) Conversation Games and a Strategic View of the Turing Test.
(Nevens et al., 2020) A Practical Guide to Studying Emergent Communication through Grounded Language Games.
(Wen et al., 31 Jan 2025) Language Games as the Pathway to Artificial Superhuman Intelligence.
(Schaul, 2024) Boundless Socratic Learning with Language Games.
(Suglia et al., 2023) Visually Grounded Language Learning: a review of language games, datasets, tasks, and models.
(Vartziotis, 1 Jan 2026) Language as Mathematical Structure: Examining Semantic Field Theory Against Language Games.
(Khani et al., 2018) Planning, Inference and Pragmatics in Sequential Language Games.
(Xu et al., 23 May 2025) CoMet: Metaphor-Driven Covert Communication for Multi-Agent Language Games.
(Landsman et al., 2023) Is mathematics a game?
(Felice et al., 2020) Functorial Language Games for Question Answering.
(Hedges et al., 2018) Towards Functorial Language-Games.
(Schlangen, 2023) Dialogue Games for Benchmarking Language Understanding: Motivation, Taxonomy, Strategy.

The language game paradigm thus offers a comprehensive lens for understanding language as a vehicle for meaning, learning, strategy, and computation—spanning formal abstraction, empirical analysis, and integrative system design.