Nameability: A Multidisciplinary Perspective

• Nameability is the property by which targets become stably and uniquely referable through shared constraints like entropy, collision risk, and context overlap.
• It is operationalized differently across fields, using metrics such as Rényi-2 entropy, salience rates, token-level precision, and cosine similarity for external validation.
• Empirical studies show that nameability affects communication efficiency, model interpretability, and privacy, underscoring its theoretical importance and practical applications.

Nameability is the property by which a target becomes linguistically referable in a stable, discriminative, and task-adequate way. Across disciplines, the term denotes closely related but non-identical phenomena: unique identification by description in a shared knowledge space, recoverability of semantic labels from low-level program artifacts, cross-speaker agreement on labels for objects, robustness of proper-name reference under ambiguity, referability of declarations under access-control policies, and the extent to which humans can accurately verbalize internal model features (Guha et al., 2015, Artuso et al., 2019, Eliav et al., 2023, Zwaan et al., 2024, Colin et al., 19 May 2026).

1. Conceptual range

Research area	What is nameable	Operational sense
Reference and information theory	An entity	Uniquely identifiable by a description
Reverse engineering	A stripped binary function	Recoverable via normalized function-name tokens
Psycholinguistics and communication	A person, character, or novel referent	Learnable, retrievable, and shared across speakers
Logic and programming languages	An agent, group, or declaration	Referable through admissible semantics or resolution paths
Model interpretability	A learned feature	Accurately describable in natural language

In information-theoretic work on reference, nameability is the property that an intended entity can be uniquely singled out, with high probability, by a description constructed from shared knowledge. In stripped-binary analysis, it is the degree to which a model can recover semantically meaningful function names from assembly alone. In distributional semantics, it concerns the learnability, retrievability, and separability of representations for individual entities, especially proper names. In communication studies on abstract stimuli, it denotes the a priori likelihood that two individuals will independently choose the same label. In formal logic and programming-language theory, it shifts from lexical labeling to referability: whether a name can rigidly or non-rigidly denote, and whether a declaration can be referred to at a program point. In interpretability research, it becomes a behavioral measure of whether humans can state what a learned feature represents (Guha et al., 2015, Artuso et al., 2019, Bruera et al., 2021, Eliav et al., 2023, Bílková et al., 2021, Zwaan et al., 2024, Colin et al., 19 May 2026).

This breadth does not make the concept vacuous. A common core is the relation between a target, a naming system, and a background of constraints. The constraints may be entropy and collision risk, low-level code normalization, context overlap among story characters, shared conventions in dyadic interaction, accessibility predicates on scope-graph paths, or human psychophysics over feature visualizations. Nameability therefore functions less as a single metric than as a family of task-specific criteria for successful reference.

2. Formalizations and metrics

The most explicit information-theoretic treatment models nameability through collision probabilities and threshold phenomena. In the feature-vector setting, if descriptors are independent, uniqueness is controlled by cumulative Rényi-2 entropy, with a sufficient condition

$$\sum_{j=1}^k H_2(X_j) \ge 2\log n + \log(1/\delta).$$

In the graph setting, if a description has $L$ arcs and salience rate $F$, a corresponding threshold is

$L \ge \frac{2\log n + \log(1/\delta)}{F}.$

That work also proposes a Nameability Index,

$$\mathcal{N} = \frac{\sum_{j=1}^k H_2(X_j)}{2\log n},$$

or, for graph descriptions, its salience-based analogue, with the interpretation that $\mathcal{N} \ge 1$ places the system in the “nameable regime” (Guha et al., 2015).

Other fields replace uniqueness with agreement, recoverability, or alignment. In stripped-binary function naming, the task is sequence transduction from normalized assembly tokens $x$ to normalized name tokens $y$, trained by token-level cross-entropy and evaluated by bag-of-tokens precision, recall, and $F1$, not by exact string match, BLEU, or ROUGE-L. The membership function

$$\alpha(\hat{x}_i, x)= \begin{cases} 1 & \text{if } \hat{x}_i \in x\ 0 & \text{else} \end{cases}$$

supports token-level

$L$0

This operationalization explicitly treats “nameability” as recoverable semantic content under severe information loss (Artuso et al., 2019).

In visual communication with tangrams, nameability is operationalized by Shape Naming Divergence (SND), the mean proportion of words in a description that do not appear in any other description of the same shape, with a corresponding index $L$1. In visual object naming more broadly, low entropy of the human label distribution indicates high nameability, and a normalized entropy-based index

$L$2

serves the same purpose. In Mandarin object naming, Shannon entropy

$L$3

is used directly as the inverse of nameability: lower $L$4 means stronger name agreement (Ji et al., 6 Sep 2025, Testoni et al., 2024, He et al., 2023).

Interpretability work introduces a different metric family. Participants write free-text descriptions of recovered vision-model features, and each response is scored by the mean cosine similarity between the text embedding and the embeddings of nine feature-centered image crops in CLIP space. The reported chance level is approximately $L$5, defined by mismatched image-text pairs from different features. Here nameability is neither uniqueness nor agreement across speakers, but externally validated semantic describability (Colin et al., 19 May 2026).

Privacy-oriented work treats nameability as context-sensitive named-entity recognizability. It uses standard precision, recall, and $L$6, and quantifies the ambiguity effect by relative recall drop

$L$7

This makes explicit that a string may be a human name and yet be weakly nameable for a detector if context and model biases push it toward a non-person interpretation (Pham et al., 20 May 2025).

3. Experimental paradigms and datasets

The empirical study of nameability depends heavily on task design. In reverse engineering, UbuntuDataset comprises 8.86 million functions from 22,040 packages, disassembled from 87,853 distinct ELF files; after filtering, names are normalized to a 1064-token vocabulary, and functions are represented as flattened instruction sequences with average length 159.12 instructions. This setting isolates semantic recoverability from assembly alone by excluding control/data-flow graphs and explicit call context (Artuso et al., 2019).

In distributional semantics, the Novel Aficionados dataset comprises 59 novels, each split into two halves. The Doppelgänger test asks whether a representation learned for an entity in one half can identify its co-referent in the other half among competing candidates. The task is deliberately difficult because characters share a narrative universe and often occur in similar contexts; this is the central setup used to compare proper names and common nouns (Bruera et al., 2021).

Dyadic communication studies use the KiloGram dataset of 1016 black-and-white tangram images. One line of work sorts items by SND and constructs high- and low-nameability conditions, then embeds them in repeated reference games with alternating speaker and listener roles. A later, larger study uses 151 dyads and a three-phase design—pre, interactive, and post—with “threads” of visually similar tangrams spanning similarity bins derived from fine-tuned CLIP image embeddings (Eliav et al., 2023, Ji et al., 6 Sep 2025).

Visual naming research uses heterogeneous resources matched to different aspects of label choice. ManyNames provides names for 25,000 objects from Visual Genome with an average of 35.3 human annotations and more than one unique label for over 90% of objects; NOUN provides 64 images of unusual multipart objects for studying attribute-heavy naming; QUANT provides 850 abstract scenes for quantifier choice over target proportions; ManyNames ZH contributes 1,319 naturalistic images with approximately 20 Mandarin names per object (Testoni et al., 2024, He et al., 2023).

The privacy benchmark AMBENCH is constructed specifically to stress-test the nameability of ambiguous human names. It curates approximately 12,000 human names confusable with locations, organizations, syndromes, bacteria, and minerals, places them into roughly 60,000 short English snippets, and adds benign prompt injections that resemble ordinary user instructions. The benchmark also includes a 10,000-snippet baseline built from the top-200 US first names (Pham et al., 20 May 2025).

Human-interpretability work on vision models contributes a further paradigm. Features are recovered by TopK sparse autoencoders trained on ImageNet-1k patch-token activations; the study analyzes more than 15,000 behavioral responses, of which 13,400 from 377 participants pass pre-specified quality checks. Nameability trials combine synthetic visualizations, top-activating images, and RISE heatmaps, thereby making feature verbalization a controlled psychophysics task rather than a post hoc anecdotal judgment (Colin et al., 19 May 2026).

4. Empirical regularities

A recurring empirical pattern is that nameability is systematically easier for some target classes than for others. In the Doppelgänger test, proper names are less clearly distinguishable from each other than common nouns across ELMo, BERT, Word2Vec, and Nonce2Vec, and the reported explanation is competition among similar individual entities occupying overlapping contexts (Bruera et al., 2021). In Mandarin object naming, familiarity has an overall negative coefficient on entropy, $L$8 with 95% credible interval $L$9, indicating that more familiar items tend to elicit more convergent naming, but PEOPLE shows the highest predicted variation while ANIMALS_PLANTS shows the lowest (He et al., 2023).

In dyadic communication, higher nameability improves immediate coordination. In the 151-dyad tangram study, higher nameability increased accuracy with $F$0, 95% CI $F$1, shortened descriptions with $F$2, 95% CI $F$3, and reduced response time with $F$4, 95% CI $F$5. Yet the same study reports that pre–post alignment for undiscussed tangrams increased overall and that nameability did not significantly modulate the pre–post change, $F$6, 95% CI $F$7, so convention generalization was “robust across levels of nameability” (Ji et al., 6 Sep 2025). A related two-experiment study found that transfer to entirely new targets and contexts was stronger for high-nameability tangrams, including a significant train-vs-test interaction $F$8, $F$9 (Eliav et al., 2023).

In stripped binaries, nameability is feasible but imperfect under strong information constraints. On UbuntuDataset test data, Seq2Seq reaches precision $L \ge \frac{2\log n + \log(1/\delta)}{F}.$0, recall $L \ge \frac{2\log n + \log(1/\delta)}{F}.$1, and $L \ge \frac{2\log n + \log(1/\delta)}{F}.$2 $L \ge \frac{2\log n + \log(1/\delta)}{F}.$3, whereas the Transformer reaches precision $L \ge \frac{2\log n + \log(1/\delta)}{F}.$4, recall $L \ge \frac{2\log n + \log(1/\delta)}{F}.$5, and $L \ge \frac{2\log n + \log(1/\delta)}{F}.$6 $L \ge \frac{2\log n + \log(1/\delta)}{F}.$7. Fine-tuning is especially consequential under domain shift: on the malware dataset, average family-level $L \ge \frac{2\log n + \log(1/\delta)}{F}.$8 rises from approximately $L \ge \frac{2\log n + \log(1/\delta)}{F}.$9 for the pretrained Transformer to approximately $$\mathcal{N} = \frac{\sum_{j=1}^k H_2(X_j)}{2\log n},$$0 after fine-tuning (Artuso et al., 2019).

Vision-and-LLMs exhibit mixed nameability. For common objects in ManyNames, LLaVA reaches inverse JSD around $$\mathcal{N} = \frac{\sum_{j=1}^k H_2(X_j)}{2\log n},$$1 and outperforms FROMAGe and BLIP-2, but none of the models significantly correlates with human texture saliency on NOUN, and overall Pearson’s $$\mathcal{N} = \frac{\sum_{j=1}^k H_2(X_j)}{2\log n},$$2 on QUANT is close to zero, indicating failure on graded quantifier choice (Testoni et al., 2024). Human interpretability studies of vision foundation models show a parallel pattern: median nameability scores are $$\mathcal{N} = \frac{\sum_{j=1}^k H_2(X_j)}{2\log n},$$3 for ViT-S/16 and $$\mathcal{N} = \frac{\sum_{j=1}^k H_2(X_j)}{2\log n},$$4 for ViT-B/16, while CLIP, DINOv3, DINOv2, and SigLIP score $$\mathcal{N} = \frac{\sum_{j=1}^k H_2(X_j)}{2\log n},$$5, $$\mathcal{N} = \frac{\sum_{j=1}^k H_2(X_j)}{2\log n},$$6, $$\mathcal{N} = \frac{\sum_{j=1}^k H_2(X_j)}{2\log n},$$7, and $$\mathcal{N} = \frac{\sum_{j=1}^k H_2(X_j)}{2\log n},$$8 respectively; differences are significant with Kruskal–Wallis $$\mathcal{N} = \frac{\sum_{j=1}^k H_2(X_j)}{2\log n},$$9, $\mathcal{N} \ge 1$0 (Colin et al., 19 May 2026).

Privacy tasks reveal that low nameability can be operationally dangerous. Across models and tools, ambiguous human names in AMBENCH have recall between $\mathcal{N} \ge 1$1 and $\mathcal{N} \ge 1$2, whereas regular names are typically at or above $\mathcal{N} \ge 1$3, yielding average relative recall drops in the $\mathcal{N} \ge 1$4–$\mathcal{N} \ge 1$5 band for many systems. In privacy-preserving summarization, leakage for ambiguous-name conversations rises from $\mathcal{N} \ge 1$6 without benign prompt injection to $\mathcal{N} \ge 1$7 with injection, corresponding to a risk ratio of approximately $\mathcal{N} \ge 1$8, while the regular-name baseline remains essentially flat at $\mathcal{N} \ge 1$9 (Pham et al., 20 May 2025).

5. Determinants, interventions, and applications

The strongest cross-domain determinants of nameability are discriminative signal, competition, and alignment. In reference by description, high-entropy descriptors and graph salience suppress collisions; adaptive search over candidate descriptions reduces expected description length, while correlated descriptors degrade performance because joint discriminative power grows sublinearly (Guha et al., 2015). In psycholinguistic and distributional settings, the central degraders are overlapping contexts and lack of definitional redundancy: proper names denote unique individuals but often lack the high-mutual-information co-occurrences that stabilize common-noun representations (Bruera et al., 2021).

Several interventions recur. In stripped-binary analysis, domain-adaptive fine-tuning on target domains such as optimization levels, system utilities, or malware substantially improves function nameability, which suggests that pretraining mostly supplies reusable low-level statistical regularities while domain data repairs the mapping to task-specific semantic vocabularies (Artuso et al., 2019). In VLLMs, the proposed remedies are pragmatic modeling, explicit counting or scene-graph reasoning for quantifiers, threshold learning for graded terms such as “few” and “most,” and distributional training against human label distributions rather than single gold labels (Testoni et al., 2024).

Perceptual design work makes nameability itself an optimization target. Color-Name Aware optimization represents each color as a distribution over color terms, measures name similarity by cosine similarity between those distributions, and optimizes palette, opacity, and layer order so that composite colors remain name-coherent with their parent classes while preserving discriminability. In crowdsourced evaluations on overlapped histograms, this approach yields lower error than standard, local, and hue-preserving blending baselines in both distribution estimation and class discrimination tasks (Lu et al., 2024).

Business and branding treat nameability in still another applied sense: the capacity of a candidate name to create value across markets, channels, and time. In that setting, a name must be distinctive and legally registrable, easy to pronounce and remember, semantically aligned with brand positioning, culturally safe, and findable in digital ecosystems. The broader onomastic argument is that globalization, digitalization, and Big Data turn naming from an initial branding exercise into a continuous analytic problem spanning marketing, HR, communication, and risk management (Carsenat, 2013).

Dynamic naming systems also show that communicative efficiency alone does not force a one-to-one lexicon. In the $x$0-object naming game, homonymy persists as a dynamical trap, synonymy becomes negligible in the long-time limit, and mild noise can regroup words more evenly in verbal space even while reducing communication efficiency (0810.3442). This suggests that nameability is often constrained not only by signal quality but by the update dynamics of the naming system itself.

6. Logical and semantic nameability

Formal work in logic and programming languages reinterprets nameability as admissible reference. In epistemic logic with assignments, names are non-rigid and variables are rigid; a formula such as

$x$1

expresses that agent $x$2 knows who $x$3 is by binding a variable to the actual denotation of $x$4 and then requiring that this denotation be preserved across $x$5’s epistemic alternatives. This framework is designed precisely for situations in which names are not commonly known, pseudonyms vary, or agents disagree about reference (Wang et al., 2018).

A related epistemic logic replaces fixed agents and extensional groups with intensional names $x$6 whose extensions $x$7 vary across worlds. The resulting modalities $x$8, $x$9, $y$0, and $y$1 do not rigidly name individuals; they describe de dicto epistemic attitudes of whoever currently bears the name. The system makes “the current $y$2-group” nameable, while leaving fixed-agent properties and cardinality constraints outside the expressive reach of the base language (Bílková et al., 2021).

Programming-language theory gives nameability a path-semantic reading. In scope-graph terms, a name is nameable at a reference site if there exists a resolution path to a declaration that is not shadowed and that satisfies the language’s accessibility policy. Nameability is therefore the conjunction of visibility and accessibility, formalized declaratively as a predicate on resolution paths and instantiated for Java, C#, C++, Rust, and Eiffel (Zwaan et al., 2024).

Nominal logic and coalgebraic semantics generalize the same theme. In Permissive-Nominal Logic, names are atoms, globally available and permutable; semantic nameability is controlled by support, and terms denote elements whose supports are bounded by their free atoms. In coalgebra on nominal sets, regular name-dependent behaviors are exactly the rational behaviors preserved by orbit-finite structure, localizable liftings, and quotienting by binding or $y$3-equivalence. On that view, nameability is constrained by what remains representable and recognizable under symmetry, support, and regularity (Dowek et al., 2023, Milius et al., 2016).

Taken together, these formal accounts show that nameability is not reducible to lexical choice. It can denote stable reference under epistemic uncertainty, legal referability under access control, or semantic representability under nominal symmetry. The shared theme is that a nameable object is one for which the governing system supplies a tractable route from designation to identification.