Morphological Communication Overview

• Morphological communication is a cross-disciplinary concept where structured forms actively determine how information is encoded, transmitted, and coordinated.
• It spans diverse domains including lexicalization in linguistics, morphology-mediated control in robotics, and emergent inflection in neural communication games.
• The field highlights that morphology functions as both constraint and channel, unifying computational, geometric, and material perspectives in effective information flow.

Morphological communication is a cross-disciplinary term used for communicative processes in which morphology is not merely a passive substrate but an active determinant of encoding, transmission, coordination, or decoding. In recent work, the term has at least five technically distinct uses: as a model of lexicalization in word-internal morphology, where speakers choose among competing morpheme sequences under a recoverability–cost trade-off; as morphology-mediated information exchange in embodied and biological systems; as mechanically mediated coordination in soft robots; as the emergence of morpheme-like structure in neural communication games; and as a structural or geometric account of communication codes in linguistics (Yang et al., 5 May 2026).

1. Terminological scope and core distinctions

The term is used in several non-equivalent senses across current research.

Domain	Morphological communication denotes	Representative paper
Historical lexicalization	Choice of morpheme sequences under listener recoverability and speaker production cost	(Yang et al., 5 May 2026)
ICON / morphological computing	Exchange and shaping of information through material structure and physical interaction	(Dodig-Crnkovic, 2024)
Soft robotics	A compliant body acting as an emergent “data bus” for decentralized controllers	(Meek et al., 27 Aug 2025)
Neural emergent communication	Mapping meanings to morpheme-like multi-character segments under double articulation	(Gilberti et al., 7 Aug 2025)
Language theory and geometry	Structural signal–meaning codes or geometric mappings between feature bundles and exponents	(Seoane et al., 2018, Goldsmith et al., 2017)

A recurrent source of confusion is the relation between morphological communication and morphological computation. In the ICON / MC literature, morphological computation is the transformation of information by material dynamics, whereas morphological communication is the propagation or coordination of information through morphology-mediated coupling; the two are treated as complementary rather than identical (Dodig-Crnkovic, 2024). A closely related distinction appears in soft robotics, where morphological computation concerns useful physical transformations of control signals, while morphological communication concerns the transmission of information among controller modules through shared body dynamics (Meek et al., 27 Aug 2025).

Another recurring ambiguity concerns the word morphology itself. In linguistics, it refers to word structure and inflectional or derivational composition. In robotics and embodied cognition, it refers to bodily form, materials, compliance, and dynamical constraints. The literature surveyed here uses the same term for both, but the communicative mechanisms differ: morpheme selection in one case, morphodynamic coupling in the other. A plausible implication is that the phrase functions less as a single theory than as a family of theories unified by the claim that structured form shapes communication.

2. Word-internal communication and lexicalization

In historical lexical semantics and morphology, morphological communication has been formalized as the selection of a morpheme sequence for a target concept from a set of contemporaneously available alternatives (Yang et al., 5 May 2026). The central claim is that lexicalization reflects a communicative trade-off between listener recoverability and speaker efficiency. Morphological composition is defined as the combinatorial assembly of morphemes into new words; compounding combines free morphemes or lexemes, while derivation or affixation combines a base with bound morphemes. Lexicalization is the conventionalization of a form–meaning pairing in the lexicon, that is, the process by which a language “settles” on one morpheme combination among alternatives for a concept.

The formalism is cast in the Rational Speech Act framework. A morpheme sequence $u = (\mu_1,\dots,\mu_m)$ names a concept $c$ within a time-indexed lexicon $\mathcal{L}_t$. Listener recoverability is operationalized through semantic compatibility and a literal listener:

$$L_0(c \mid u, \mathcal{L}_t) \propto S_0(u \mid c, \mathcal{L}_t)\, P(c \mid \mathcal{L}_t).$$

Speaker production cost is additive over morphemes:

$$\mathrm{Cost}(u,\mathcal{L}_t)=\sum_{i=1}^{m} h_\phi\big(\mathrm{feat}(\mu_i,\mathcal{L}_t)\big).$$

The pragmatic speaker then trades informativeness against cost:

$$S_1(u \mid c,\mathcal{L}_t)\propto \exp\big(\log L_0(c \mid u,\mathcal{L}_t)-\mathrm{Cost}(u,\mathcal{L}_t)\big).$$

The implemented linear utility is

$$U_\theta(u,c)=\alpha \cdot \tilde{f}_\theta(u,c)-\beta \cdot \mathrm{Cost}(u,\mathcal{L}_t)+b,$$

with a nonlinear alternative using an $\mathrm{MLP}_\theta$ over the same two components.

The empirical setup uses a time-indexed lexicon built from COHA and COCA over 1820–2019, with emergence time defined as the first year with non-zero counts and candidate alternatives constructed from morphemes available at that time. The study evaluates 4323 naturally occurring English compounds and derivations from WordNet, with glosses, MorphSeg segmentations, and pronunciation features from the CMU Pronouncing Dictionary. Candidate sets are sampled up to 1024 per item, and evaluation uses Mean Reciprocal Rank and top-$k$ accuracy:

$$\mathrm{MRR}=\frac{1}{N}\sum_{i=1}^{N}\frac{1}{r_i}, \qquad \mathrm{Acc@}k=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}[r_i\leq k].$$

The reported ranking results show a consistent ordering among interpretable models:

Model	MRR	Brief characterization
Cost-only	0.031	Weakest
Semantic-only	0.047	Meaning helps
S1 (linear/nonlinear)	0.050–0.053	Modest but consistent gain
Discriminative	0.096	Best overall

At concrete thresholds, Nonlinear $c$0 reaches $c$1 and $c$2, compared with $c$3 and $c$4 for the semantic-only baseline. The advantage of $c$5 over semantic-only grows as the candidate set expands, which is interpreted as evidence that meaning alone leaves morphological choice underdetermined. Averaged fitted weights in the linear model suggest that both pressures are active, with semantic weight $c$6, coefficient on the negated cost-based score $c$7, and $c$8.

Qualitative cases illustrate the failure modes of one-factor accounts. For laundry ($c$9), $\mathcal{L}_t$0 ranks the attested composition first, while cost-only prefers short frequent fragments such as ed+ing and semantic-only admits longer meaning-adjacent variants such as laundr + ify + ing. For cynicism ($\mathcal{L}_t$1), semantic-only favors redundant or ill-formed alternatives such as cynic + cynic and cynic + antipathy, whereas $\mathcal{L}_t$2 penalizes redundancy and morphological excess. These cases are used to argue that attested composition is shaped by a joint pressure toward expressiveness and economy rather than by structural rule application or semantic compatibility alone.

The framework is explicitly limited. Candidates are treated as order-agnostic; morphotactic constraints and head–modifier structure are not explicitly modeled; semantic scoring uses contemporary Qwen3 embeddings while diachronic change is represented only indirectly through historical word2vec neighbors and frequency features. The performance gap between the RSA-style $\mathcal{L}_t$3 models and the stronger discriminative model is taken to indicate that richer feature interactions remain unmodeled (Yang et al., 5 May 2026).

3. Embodied, observer-relative, and morphodynamic communication

In the info-computational and morphological computing literature, morphological communication is the exchange and shaping of information through material structure and physical interaction (Dodig-Crnkovic, 2024). ICON treats natural structures as information and processes of change in those structures as computation; MC focuses on computations realized by morphology. The framework is explicitly relational and observer-relative: what counts as informational structure and which material transformations implement computation depend on the perspective of a cognizer. Landauer’s principle is adopted in the strong form that information is always tied to a physical representation.

Within this framework, morphology is informational because shape, stiffness, topology, and chemical gradients make a difference to downstream interactions and predictions. Morphological computation is the process by which such morphological change implements information processing, including morphology facilitating control, morphology facilitating perception, and morphological computation proper. Morphological communication denotes the transmission and coordination of information through morphology-mediated physical interaction. The same morphodynamic process may therefore both compute, by changing informational state, and communicate, by making that state available to other parts or agents through physical coupling.

The chapter situates this account within the Free Energy Principle and active inference. Variational free energy is given in the cited literature as

$\mathcal{L}_t$4

and expected free energy for a policy $\mathcal{L}_t$5 as

$\mathcal{L}_t$6

Perception and action are described as gradient flows on free energy,

$\mathcal{L}_t$7

where morphology affects the likelihood $\mathcal{L}_t$8 through sensing and body–environment contact, and affects dynamics through passive mechanics. Morphogenesis is correspondingly interpreted as Bayesian inference: cells and tissues minimize a free-energy-like quantity over bioelectric, chemical, and mechanical fields, with morphology both computing and communicating the evolving pattern.

The chapter also reviews information-theoretic measures aligned with this perspective. Mutual information,

$\mathcal{L}_t$9

quantifies shared information between body or morphology variables and outcomes. Transfer entropy,

$$L_0(c \mid u, \mathcal{L}_t) \propto S_0(u \mid c, \mathcal{L}_t)\, P(c \mid \mathcal{L}_t).$$0

captures directed information flow. A typical measure of body contribution is

$$L_0(c \mid u, \mathcal{L}_t) \propto S_0(u \mid c, \mathcal{L}_t)\, P(c \mid \mathcal{L}_t).$$1

which isolates how much of the next state is due to morphology and dynamics beyond direct control. The chapter further gives a dynamical systems formalization in which internal states $$L_0(c \mid u, \mathcal{L}_t) \propto S_0(u \mid c, \mathcal{L}_t)\, P(c \mid \mathcal{L}_t).$$2, morphology $$L_0(c \mid u, \mathcal{L}_t) \propto S_0(u \mid c, \mathcal{L}_t)\, P(c \mid \mathcal{L}_t).$$3, environment $$L_0(c \mid u, \mathcal{L}_t) \propto S_0(u \mid c, \mathcal{L}_t)\, P(c \mid \mathcal{L}_t).$$4, and observations $$L_0(c \mid u, \mathcal{L}_t) \propto S_0(u \mid c, \mathcal{L}_t)\, P(c \mid \mathcal{L}_t).$$5 are coupled through functions $$L_0(c \mid u, \mathcal{L}_t) \propto S_0(u \mid c, \mathcal{L}_t)\, P(c \mid \mathcal{L}_t).$$6; morphological communication is encoded in the coupling terms $$L_0(c \mid u, \mathcal{L}_t) \propto S_0(u \mid c, \mathcal{L}_t)\, P(c \mid \mathcal{L}_t).$$7, while morphological computation is encoded in the state transformations themselves.

Examples are drawn across scales. In single cells, quorum sensing broadcasts population state through extracellular concentration fields. In morphogenesis, reaction–diffusion, adhesion, bioelectrical signaling, and mechanical feedback perform non-symbolic computation while tissue deformation and gradients communicate the result to neighboring cells. In whole organisms, passive dynamics and compliance both pre-process sensory signals and coordinate movement through contact forces and posture. In collectives, stigmergy and niche construction are treated as cases in which environmental morphology communicates constraints and opportunities across time. In robotics, compliant bodies and passive dynamics reduce control complexity while physical coupling distributes information without centralized symbols.

A plausible implication of this framework is that communication can be materially distributed without explicit symbols or packets. The chapter nonetheless emphasizes open problems: standardizing definitions across domains, integrating information-theoretic metrics with FEP under observer-relative assumptions, and performing perturbation-based experimental validation in tissues, plants, soft bodies, and swarms (Dodig-Crnkovic, 2024).

4. Soft robotic bodies as emergent data buses

In soft robotics, morphological communication has been defined more narrowly as the use of a robot’s compliant, dynamically coupled body as an emergent “data bus” through which independent controller modules coordinate without explicit electronic signaling (Meek et al., 27 Aug 2025). The setting is a simulated soft robot in EvoGym. Each actuator has its own local controller and no direct knowledge of the others; inputs are limited to body-centric telemetry. When one actuator changes its target state, the resulting deformation propagates through shared springs and point masses, thereby altering the sensory signals seen by other controllers. Coordination is therefore mechanically mediated rather than symbolically transmitted.

The body is modeled as a lattice of point masses and springs with actuator-imposed target lengths. A canonical element-wise coupling is written as

$$L_0(c \mid u, \mathcal{L}_t) \propto S_0(u \mid c, \mathcal{L}_t)\, P(c \mid \mathcal{L}_t).$$8

where $$L_0(c \mid u, \mathcal{L}_t) \propto S_0(u \mid c, \mathcal{L}_t)\, P(c \mid \mathcal{L}_t).$$9 and $$\mathrm{Cost}(u,\mathcal{L}_t)=\sum_{i=1}^{m} h_\phi\big(\mathrm{feat}(\mu_i,\mathcal{L}_t)\big).$$0 encode stiffness and damping and $$\mathrm{Cost}(u,\mathcal{L}_t)=\sum_{i=1}^{m} h_\phi\big(\mathrm{feat}(\mu_i,\mathcal{L}_t)\big).$$1 captures actuator forces induced by target-length commands. The crucial property is shared structure: adjacent voxels share point masses, so a perturbation in one actuator’s target length changes the distances sensed by others. The study argues that timing is critical because actuators approach targets over roughly 10–12 simulation steps rather than instantaneously. Sampling SNN inputs every 12 steps aligns neural updates with body transients and is reported to prevent cancellation and stabilize rhythmic patterns.

The robot platform is the EvoGym “Walk the Robot” task. Voxels are encoded in JSON as empty, rigid, soft, horizontal actuator, vertical actuator, and fixed. The exemplar “Ubot” morphology is a two-legged soft robot with six horizontal actuators and two vertical actuators, for 8 actuators and 32 point masses. Each actuator’s controller receives two distances: its center of mass to the robot’s global top-left and bottom-right corners. Because all controllers reference the same global corners, deformations anywhere in the body influence all sensors.

Control is decentralized. Each actuator has an independent spiking neural network with 2 inputs, 2 hidden leaky integrate-and-fire neurons, and 1 output neuron. There is no synaptic coupling across actuators. In continuous-time form, the cited neuron model is

$$\mathrm{Cost}(u,\mathcal{L}_t)=\sum_{i=1}^{m} h_\phi\big(\mathrm{feat}(\mu_i,\mathcal{L}_t)\big).$$2

with synaptic current conceptually given by

$$\mathrm{Cost}(u,\mathcal{L}_t)=\sum_{i=1}^{m} h_\phi\big(\mathrm{feat}(\mu_i,\mathcal{L}_t)\big).$$3

Output spikes are decoded into target lengths via

$$\mathrm{Cost}(u,\mathcal{L}_t)=\sum_{i=1}^{m} h_\phi\big(\mathrm{feat}(\mu_i,\mathcal{L}_t)\big).$$4

with $$\mathrm{Cost}(u,\mathcal{L}_t)=\sum_{i=1}^{m} h_\phi\big(\mathrm{feat}(\mu_i,\mathcal{L}_t)\big).$$5. The study also explored an ad hoc STDP-like heuristic, but reports that it was unstable and was not used in the main experiments.

Optimization uses CMA-ES over a flat parameter vector of length 72, corresponding to 8 actuators times 9 parameters per SNN. Hyperparameters are population size 12, initial mean the zero vector in $$\mathrm{Cost}(u,\mathcal{L}_t)=\sum_{i=1}^{m} h_\phi\big(\mathrm{feat}(\mu_i,\mathcal{L}_t)\big).$$6, and initial sigma 1. Fitness is average forward displacement along $$\mathrm{Cost}(u,\mathcal{L}_t)=\sum_{i=1}^{m} h_\phi\big(\mathrm{feat}(\mu_i,\mathcal{L}_t)\big).$$7:

$$\mathrm{Cost}(u,\mathcal{L}_t)=\sum_{i=1}^{m} h_\phi\big(\mathrm{feat}(\mu_i,\mathcal{L}_t)\big).$$8

Each simulation runs for 1000 time steps, with telemetry and actions updated every 12 steps. To avoid stagnation, some experiments evolve 30 parallel populations.

The main evidence is behavioral and qualitative. After penalizing fallen postures and tuning input scaling and sampling, robots evolved genuine multi-actuator rhythmic gaits; an exemplar traversed the platform fully after 64 generations. Fitness improved from approximately 100 initially down to 27 on the 0–100 scale, with diminishing returns after about 28 generations. Horizontal actuator modules often converged to highly similar activation patterns, while a vertical actuator exhibited offset spiking aligned to ground contact, which is interpreted as evidence that it “listens” to body state changes induced by other actuators and the environment. Without 12-step throttling, gaits reportedly degraded into vibration.

The paper is explicit that formal quantification remains future work. Suitable measures are proposed, including mutual information between actuator $$\mathrm{Cost}(u,\mathcal{L}_t)=\sum_{i=1}^{m} h_\phi\big(\mathrm{feat}(\mu_i,\mathcal{L}_t)\big).$$9’s output sequence $$S_1(u \mid c,\mathcal{L}_t)\propto \exp\big(\log L_0(c \mid u,\mathcal{L}_t)-\mathrm{Cost}(u,\mathcal{L}_t)\big).$$0 and actuator $$S_1(u \mid c,\mathcal{L}_t)\propto \exp\big(\log L_0(c \mid u,\mathcal{L}_t)-\mathrm{Cost}(u,\mathcal{L}_t)\big).$$1’s sensor stream $$S_1(u \mid c,\mathcal{L}_t)\propto \exp\big(\log L_0(c \mid u,\mathcal{L}_t)-\mathrm{Cost}(u,\mathcal{L}_t)\big).$$2,

$$S_1(u \mid c,\mathcal{L}_t)\propto \exp\big(\log L_0(c \mid u,\mathcal{L}_t)-\mathrm{Cost}(u,\mathcal{L}_t)\big).$$3

transfer entropy $$S_1(u \mid c,\mathcal{L}_t)\propto \exp\big(\log L_0(c \mid u,\mathcal{L}_t)-\mathrm{Cost}(u,\mathcal{L}_t)\big).$$4, Granger causality, phase-locking value, and dynamical sensitivity $$S_1(u \mid c,\mathcal{L}_t)\propto \exp\big(\log L_0(c \mid u,\mathcal{L}_t)-\mathrm{Cost}(u,\mathcal{L}_t)\big).$$5. The current study infers morphological communication from coordinated behavior rather than from direct information-flow analysis. That limitation is central to its interpretation (Meek et al., 27 Aug 2025).

5. Emergent inflectional morphology in neural communication games

A different use of the term appears in neural emergent communication, where communication is reformulated as the emergence of inflectional morphology under a small-vocabulary constraint (Gilberti et al., 7 Aug 2025). In this setting, agents communicate attribute–value bundles through sequences of discrete characters. Morphological communication denotes the mapping between meanings and multi-character segments functioning like morphemes. The authors explicitly target inflectional rather than derivational morphology by combining a high-cardinality “root” attribute with few-valued grammatical attributes such as tense and person.

The task is an attribute–value reconstruction game. States are tuples

$$S_1(u \mid c,\mathcal{L}_t)\propto \exp\big(\log L_0(c \mid u,\mathcal{L}_t)-\mathrm{Cost}(u,\mathcal{L}_t)\big).$$6

and the sender and receiver implement mappings $$S_1(u \mid c,\mathcal{L}_t)\propto \exp\big(\log L_0(c \mid u,\mathcal{L}_t)-\mathrm{Cost}(u,\mathcal{L}_t)\big).$$7 and $$S_1(u \mid c,\mathcal{L}_t)\propto \exp\big(\log L_0(c \mid u,\mathcal{L}_t)-\mathrm{Cost}(u,\mathcal{L}_t)\big).$$8. Messages are sequences up to length $$S_1(u \mid c,\mathcal{L}_t)\propto \exp\big(\log L_0(c \mid u,\mathcal{L}_t)-\mathrm{Cost}(u,\mathcal{L}_t)\big).$$9 over a fixed alphabet $$U_\theta(u,c)=\alpha \cdot \tilde{f}_\theta(u,c)-\beta \cdot \mathrm{Cost}(u,\mathcal{L}_t)+b,$$0, with the main experiments using $$U_\theta(u,c)=\alpha \cdot \tilde{f}_\theta(u,c)-\beta \cdot \mathrm{Cost}(u,\mathcal{L}_t)+b,$$1 and $$U_\theta(u,c)=\alpha \cdot \tilde{f}_\theta(u,c)-\beta \cdot \mathrm{Cost}(u,\mathcal{L}_t)+b,$$2. Because the number of values exceeds the alphabet size, agents cannot allocate one character per attribute–value pair. This enforces a double-articulation regime in which meaningful units must be built from combinations of meaningless units.

Training follows mixed optimization with REINFORCE for the sender and cross-entropy for the receiver. The sender objective is

$$U_\theta(u,c)=\alpha \cdot \tilde{f}_\theta(u,c)-\beta \cdot \mathrm{Cost}(u,\mathcal{L}_t)+b,$$3

with gradient

$$U_\theta(u,c)=\alpha \cdot \tilde{f}_\theta(u,c)-\beta \cdot \mathrm{Cost}(u,\mathcal{L}_t)+b,$$4

Receiver loss sums cross-entropies across attributes. In the inflectional setting, reward is weighted so that the root attribute receives $$U_\theta(u,c)=\alpha \cdot \tilde{f}_\theta(u,c)-\beta \cdot \mathrm{Cost}(u,\mathcal{L}_t)+b,$$5. The study uses single-layer GRUs with hidden dimension 500 in EGG, 8 random seeds per condition, and a success threshold of greater than 90% reconstruction accuracy.

The paper argues that standard emergent-communication metrics are insufficient for morphology, and it introduces dedicated measures for segmentation, concatenativity, and fusionality. Topographic similarity is defined as

$$U_\theta(u,c)=\alpha \cdot \tilde{f}_\theta(u,c)-\beta \cdot \mathrm{Cost}(u,\mathcal{L}_t)+b,$$6

Bag-of-symbols disentanglement is computed over either raw characters or segmented units, and the ratio

$$U_\theta(u,c)=\alpha \cdot \tilde{f}_\theta(u,c)-\beta \cdot \mathrm{Cost}(u,\mathcal{L}_t)+b,$$7

tests whether segmented symbols carry more attribute-specific information than characters. Concatenativity is operationalized by the average number of segments per message under HAS or BPE segmentation:

$$U_\theta(u,c)=\alpha \cdot \tilde{f}_\theta(u,c)-\beta \cdot \mathrm{Cost}(u,\mathcal{L}_t)+b,$$8

Fusionality is measured with fused TopSim:

$$U_\theta(u,c)=\alpha \cdot \tilde{f}_\theta(u,c)-\beta \cdot \mathrm{Cost}(u,\mathcal{L}_t)+b,$$9

where $\mathrm{MLP}_\theta$0 is the maximum TopSim obtained after fusing an attribute pair into a composite feature.

A further pressure is added through a toy articulation penalty that disfavors adjacent characters of identical parity:

$\mathrm{MLP}_\theta$1

with reported experiments using $\mathrm{MLP}_\theta$2. The point of this construction is not phonological realism but a simulated pronounceability constraint operating below the morpheme level.

The results support three main claims. First, double articulation emerges under the small-vocabulary constraint: in the default setting, 5 of 7 successful languages and, in the inflectional setting, all successful languages had $\mathrm{MLP}_\theta$3 under $\mathrm{MLP}_\theta$4. Second, phonological constraints encourage concatenativity. Mean $\mathrm{MLP}_\theta$5 decreases significantly with the parity-based articulation penalty, with $\mathrm{MLP}_\theta$6 in the default $\mathrm{MLP}_\theta$7 setting and $\mathrm{MLP}_\theta$8 in the inflectional $\mathrm{MLP}_\theta$9 setting. Third, fusionality patterns resemble natural-language tendencies: in 6 of 8 emergent languages in the inflectional setting, the most fused pair is tense–person rather than root–tense, while Arabic sublanguages favor tense–person fusion at approximately 78% and Spanish shows more variable behavior, with root–tense fusion dominating in approximately 48% of sublanguages and tense–person at approximately 30%.

The same study emphasizes several limitations. HAS and BPE are concatenative segmenters and perform poorly on nonconcatenative systems; up to 76% of Spanish sublanguages yield $k$0, but only up to 8% of Arabic sublanguages do so. TopSim may decrease slightly under articulation pressure, including a negative-TopSim outlier, which shows that concatenativity and global compositionality are distinct properties. Training also becomes unstable as symbolic complexity grows. The broader conclusion is that morphological structure in emergent languages is jointly shaped by phonological constraints, channel capacity, and the combinatorics of attribute systems rather than by compositionality alone (Gilberti et al., 7 Aug 2025).

6. Structural and geometric formalisms

Two further lines of research treat morphological communication as a property of formal code structure rather than as a process of lexical choice or material coupling.

The first is the morphospace approach to language networks (Seoane et al., 2018). Communication is represented by a bipartite signal–meaning matrix $k$1 with signals $k$2 and meanings $k$3. Polysemy and synonymy are encoded directly in the degrees

$k$4

Assuming meanings are equally likely and synonyms are chosen uniformly, the framework defines a speaker cost $k$5 and a hearer cost $k$6, combined by

$k$7

For synonym-free Pareto-optimal codes, the front becomes linear:

$k$8

and for $k$9 this simplifies to $$\mathrm{MRR}=\frac{1}{N}\sum_{i=1}^{N}\frac{1}{r_i}, \qquad \mathrm{Acc@}k=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}[r_i\leq k].$$0. The resulting morphospace is triangular, with the one-to-one map and the star code as extremes and a first-order phase transition at $$\mathrm{MRR}=\frac{1}{N}\sum_{i=1}^{N}\frac{1}{r_i}, \qquad \mathrm{Acc@}k=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}[r_i\leq k].$$1.

Within this framework, morphological communication is the effect of code morphology on efficiency, ambiguity, redundancy, and navigability. Polysemy raises hearer ambiguity; synonymy changes signal entropy and redundancy; component structure in the induced graphs shapes navigability. English WordNet data without particles maps near the one-to-one region and remains non-Pareto-optimal because synonymy is present in all categories. Adding a single particle linked to every meaning in nouns or verbs moves the empirical code toward the middle of the Pareto front. The paper also reports that Zipf’s law appears across a broad strip of the morphospace and does not correlate uniquely with least-effort optimization, which functions as a caution against over-identifying Zipfian statistics with one specific communicative principle.

The second formalism is geometrical morphology (Goldsmith et al., 2017), which recasts inflectional morphology as a communication channel between morphosyntactic requests and exponent vectors. A requested feature bundle is a corner $$\mathrm{MRR}=\frac{1}{N}\sum_{i=1}^{N}\frac{1}{r_i}, \qquad \mathrm{Acc@}k=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}[r_i\leq k].$$2 of a hypercube in feature-value space, while each morpheme is a unit-length vector $$\mathrm{MRR}=\frac{1}{N}\sum_{i=1}^{N}\frac{1}{r_i}, \qquad \mathrm{Acc@}k=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}[r_i\leq k].$$3 in that space. Production selects the morpheme or set of morphemes whose vector sum is closest to the requested corner:

$$\mathrm{MRR}=\frac{1}{N}\sum_{i=1}^{N}\frac{1}{r_i}, \qquad \mathrm{Acc@}k=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}[r_i\leq k].$$4

or for multiple exponents,

$$\mathrm{MRR}=\frac{1}{N}\sum_{i=1}^{N}\frac{1}{r_i}, \qquad \mathrm{Acc@}k=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}[r_i\leq k].$$5

Equivalently, morphology minimizes

$$\mathrm{MRR}=\frac{1}{N}\sum_{i=1}^{N}\frac{1}{r_i}, \qquad \mathrm{Acc@}k=\frac{1}{N}\sum_{i=1}^{N}\mathbf{1}[r_i\leq k].$$6

Comprehension inverts the same geometry by decoding the observed sum to the nearest legal corner.

This formulation makes syncretism, blocking, cumulative exponence, and portmanteaux geometric facts. A single vector can be closest to multiple corners, producing syncretism. More specific vectors win by maximizing inner product, yielding blocking. Multiple exponents sum to target complex bundles. The paper also gives data-driven learning procedures, including smart initialization from paradigm counts and a Delta Rule update with renormalization, and models inflection-class variation as rotations of a base vector configuration. In this sense, morphological communication is a high-fidelity mapping from discrete grammatical intent to continuous exponent structure.

Taken together, these two formalisms show that morphological communication can be studied at very different levels of abstraction. In one case, it is the large-scale structure of a signal–meaning code; in the other, it is the geometric alignment between feature bundles and morphological exponents. Both make morphology constitutive of communication rather than merely descriptive of it.

7. Recurrent themes, misconceptions, and open problems

Several cross-cutting themes recur despite the heterogeneity of the literature. One is the rejection of meaning-only or control-only explanations. In historical lexicalization, semantic compatibility alone does not determine which morpheme sequence becomes attested, and cost-only models over-prefer short frequent fragments (Yang et al., 5 May 2026). In soft robotics, local controllers do not require explicit message passing if body dynamics already transmit useful signals (Meek et al., 27 Aug 2025). In emergent inflection, global compositionality metrics do not suffice to characterize morphological organization because concatenativity and fusionality obey partially independent pressures (Gilberti et al., 7 Aug 2025).

A second theme is that morphology often functions as both constraint and channel. In ICON / MC, morphology carries information as physical structure and computes by changing that structure; communication and computation are two views of the same morphodynamic process (Dodig-Crnkovic, 2024). In geometrical morphology, the exponent inventory constrains what can be communicated about a feature bundle, but additive composition also provides redundancy and an error-correction-like effect (Goldsmith et al., 2017). In the morphospace model, high-degree particles can rewire the communicative code with minimal lexical change (Seoane et al., 2018).

Several misconceptions are explicitly addressed in the source literature. Morphological communication is not synonymous with morphological computation. It is not confined to human linguistic morphology. It is not reducible to explicit symbolic messaging. It does not imply that Zipf’s law uniquely diagnoses least-effort optimization. Nor does it imply that standard segmentation or compositionality metrics will recover all relevant morphological structure, especially in nonconcatenative systems.

Open problems are correspondingly domain-specific. Cross-linguistic extension is necessary for the RSA account of morphological composition, especially for agglutinative and fusional languages and for models with explicit morphotactics (Yang et al., 5 May 2026). The ICON / MC literature calls for formal unification of information-theoretic metrics, FEP, and observer-relative semantics, alongside perturbation studies in biological and robotic systems (Dodig-Crnkovic, 2024). Soft-robot studies require direct quantification of morphology-mediated information flow rather than behavioral inference alone (Meek et al., 27 Aug 2025). Emergent-communication research needs richer phonological constraints, nonconcatenative segmentation methods, and improved training stability under higher symbolic complexity (Gilberti et al., 7 Aug 2025). Network and geometric models require richer priors, weighted relations, larger systems, and better handling of grammar, noise, and sequential context (Seoane et al., 2018, Goldsmith et al., 2017).

A plausible synthesis is that morphological communication names a broad research program rather than a single doctrine. Across linguistics, cognitive science, network theory, and robotics, the common claim is that structured form—whether morpheme inventories, feature geometries, bodily materials, or signal–meaning graphs—actively determines what can be communicated, how efficiently it can be communicated, and how communicative systems evolve.