Open Zero-Shot Learning (OZSL)

• Open Zero-Shot Learning (OZSL) is a paradigm that extends traditional zero-shot learning into an open-world setting using three disjoint class sets: seen, unseen, and unknown.
• It leverages joint semantic transfer, calibration, and decision-making, integrating generative and adversarial techniques to synthesize unknown features and proxy embeddings.
• Evaluation in OZSL uses tailored metrics (e.g., F1 scores and AUROC) to measure the balance between class recognition and unknown rejection, highlighting open-world calibration challenges.

Open Zero-Shot Learning (OZSL) extends generalized zero-shot learning into an open-world regime in which a model must recognize seen classes, recognize semantically described unseen classes, and reject instances from unknown categories for which neither visual data nor class embeddings are available during training (Marmoreo et al., 2021). A closely related body of work studies open-world zero-shot recognition through enlarged candidate label spaces, especially in compositional settings where prediction must range over the full product space of primitives and many candidates are infeasible rather than merely unseen (Mancini et al., 2021). In both senses, OZSL departs from the closed-label assumption of standard ZSL and GZSL, and the central problem becomes joint semantic transfer, calibration, and open-world decision-making rather than unseen-class recognition alone (Li et al., 2023).

1. Formal task definition

The canonical OZSL formulation introduces three disjoint class sets: seen classes, unseen classes, and unknown classes. In the notation of ZS-OSR, these are

$$\mathcal{Y}^{seen},\qquad \mathcal{Y}^{unseen},\qquad \mathcal{Y}^{unknown},$$

with

$$\mathcal{Y}^{seen} \cap \mathcal{Y}^{unseen} \cap \mathcal{Y}^{unknown} = \emptyset.$$

For seen classes, training provides image samples, class labels, and semantic embeddings; for unseen classes, training provides no image samples but semantic embeddings are available; for unknown classes, training provides no image samples and no semantic embeddings (Li et al., 2023).

A standard formalization writes the seen training set as

$$\mathcal{D}^{seen}=\{(\mathbf{x}, y_{\mathbf{x}}, \mathbf{a}_{y}) \mid \mathbf{x} \in \mathcal{X}^{seen}, y_{\mathbf{x}} \in \mathcal{Y}^{seen}, \mathbf{a}_{y} \in \mathcal{A}^{seen}\},$$

and the unseen semantic inventory as

$$\mathcal{D}^{unseen}=\{(\tilde{y}_i, \mathbf{a}_{\tilde{y}_i}) \mid \tilde{y}_i \in \mathcal{Y}^{unseen}, \mathbf{a}_{\tilde{y}_i} \in \mathcal{A}^{unseen}\}.$$

At test time, data come from both unseen and unknown classes,

$\mathcal{X}^{unseen} \cup \mathcal{X}^{unknown},$

and the target is an open predictor

$$\phi^{open} : \mathbf{x} \mapsto \mathcal{O}, \qquad \mathcal{O} = \mathcal{Y}^{unseen} \cup \{\text{'unknown'}\}.$$

This is the standard unseen-versus-unknown version of OZSL; generalized versions additionally include seen classes at inference (Li et al., 2023).

The formulation introduced specifically under the name OZSL uses the class subsets

$\mathcal{S},\qquad \mathcal{U},\qquad \Omega,$

for seen, unseen, and unknown classes, respectively. The defining distinction is that seen classes have visual data and class embeddings, unseen classes have class embeddings but no visual data, and unknown classes have neither visual data nor class embeddings (Marmoreo et al., 2021). This is the sense in which the problem is “learning without seeing nor knowing”: the unknown branch requires rejection of classes the system has neither seen visually nor known semantically.

2. Relation to adjacent paradigms

OZSL is best understood relative to several neighboring formulations. Standard ZSL assumes test samples belong only to semantically specified unseen classes. GZSL allows test samples from seen and unseen classes but still assumes a closed candidate universe over those classes. Open-set recognition handles rejection of unknown classes, but does not provide zero-shot transfer to semantically specified unseen classes. OW-CZSL enlarges the candidate space to the full combinatorial product of primitives, so openness appears as semantic search over many implausible compositions rather than as a pure reject problem (Marmoreo et al., 2021).

Setting	Training assumptions	Test-time behavior
Standard ZSL	seen images; unseen semantics known	predict only unseen classes
GZSL	seen images; seen and unseen semantics known	predict over seen $\cup$ unseen
OZSL / ZS-OSR	seen images; unseen semantics known; unknowns have no samples and no semantics	classify unseen and reject unknowns; generalized variants also include seen
OW-CZSL	known primitives; seen compositions only	predict over full $\mathcal{C} = \mathcal{S} \times \mathcal{O}$
Vocabulary-informed open set	labeled source classes plus a large unlabeled semantic vocabulary	rank over a large candidate vocabulary $\mathcal{W}$

Several influential ZSL methods are explicitly not OZSL in the strict sense. OntoZSL is a standard ZSL/GZSL framework with richer ontology-based semantics but no modeling of unknown/open classes beyond a predefined unseen label set (Geng et al., 2021). ZeroMamba targets conventional ZSL and GZSL with no explicit unknown rejection (Hou et al., 2024). “Zero-Shot Learning posed as a Missing Data Problem” is a closed-set transductive unseen-only method that forces every test sample into one of the predefined unseen Gaussian components (Zhao et al., 2016). “Classifier and Exemplar Synthesis for Zero-Shot Learning” synthesizes classifiers or exemplars for a predefined unseen set and addresses seen-vs-unseen bias in GZSL, not unknown-class rejection (Changpinyo et al., 2018).

3. Evaluation protocols and metrics

A central contribution of the OZSL literature is that GZSL metrics are insufficient once unknown classes are admitted. In GZSL, the standard summary is the harmonic mean of seen and unseen recalls,

$$\mathcal{Y}^{seen} \cap \mathcal{Y}^{unseen} \cap \mathcal{Y}^{unknown} = \emptyset.$$0

but this ignores false positives from unknown samples. OZSL therefore introduces

$$\mathcal{Y}^{seen} \cap \mathcal{Y}^{unseen} \cap \mathcal{Y}^{unknown} = \emptyset.$$1

with

$$\mathcal{Y}^{seen} \cap \mathcal{Y}^{unseen} \cap \mathcal{Y}^{unknown} = \emptyset.$$2

and

$$\mathcal{Y}^{seen} \cap \mathcal{Y}^{unseen} \cap \mathcal{Y}^{unknown} = \emptyset.$$3

The motivation is explicit: unknown classes lack semantic identities, so rejection must be reflected indirectly through the false positives they induce on seen and unseen labels (Marmoreo et al., 2021).

The benchmark protocol of the original OZSL paper adapts the Proposed Splits of Xian et al. and creates unknown classes by sampling 50% of the original unseen classes and relabeling them as unknown; the remaining 50% stay unseen. Additional split variants include “20-80”, “unseen”, and “80-20” (Marmoreo et al., 2021). The same paper also reports rejection-oriented diagnostics by collapsing all unknown classes into a macro unknown bin and measuring $$\mathcal{Y}^{seen} \cap \mathcal{Y}^{unseen} \cap \mathcal{Y}^{unknown} = \emptyset.$$4, $$\mathcal{Y}^{seen} \cap \mathcal{Y}^{unseen} \cap \mathcal{Y}^{unknown} = \emptyset.$$5, and $$\mathcal{Y}^{seen} \cap \mathcal{Y}^{unseen} \cap \mathcal{Y}^{unknown} = \emptyset.$$6.

ZS-OSR work emphasizes a complementary metric suite: classification accuracy on unseen classes, $$\mathcal{Y}^{seen} \cap \mathcal{Y}^{unseen} \cap \mathcal{Y}^{unknown} = \emptyset.$$7, and $$\mathcal{Y}^{seen} \cap \mathcal{Y}^{unseen} \cap \mathcal{Y}^{unknown} = \emptyset.$$8 for unknown-vs-unseen discrimination. It also evaluates generalized ZS-OSR, ZS-OOD, and varying openness, where openness is defined as

$$\mathcal{Y}^{seen} \cap \mathcal{Y}^{unseen} \cap \mathcal{Y}^{unknown} = \emptyset.$$9

Higher openness means relatively more unknown classes (Li et al., 2023).

Compositional open-world evaluation uses another family of metrics. For OW-CZSL, the literature reports state accuracy, object accuracy, best seen accuracy, best unseen accuracy, best harmonic mean, and $$\mathcal{D}^{seen}=\{(\mathbf{x}, y_{\mathbf{x}}, \mathbf{a}_{y}) \mid \mathbf{x} \in \mathcal{X}^{seen}, y_{\mathbf{x}} \in \mathcal{Y}^{seen}, \mathbf{a}_{y} \in \mathcal{A}^{seen}\},$$0 under varying bias factors favoring unseen compositions (Mancini et al., 2021). Later OW-CZSL work preserves the seen/unseen harmonic mean and $$\mathcal{D}^{seen}=\{(\mathbf{x}, y_{\mathbf{x}}, \mathbf{a}_{y}) \mid \mathbf{x} \in \mathcal{X}^{seen}, y_{\mathbf{x}} \in \mathcal{Y}^{seen}, \mathbf{a}_{y} \in \mathcal{A}^{seen}\},$$1 emphasis, but applies them over the full compositional search space rather than a curated set of feasible labels (Kim et al., 16 May 2025).

4. Unknown-aware generative and adversarial methods

The first OZSL-specific methodological proposal is unknown feature generation. Direct Unknown Generation (DUG) synthesizes unknown visual features by MixUp-style interpolation: $$\mathcal{D}^{seen}=\{(\mathbf{x}, y_{\mathbf{x}}, \mathbf{a}_{y}) \mid \mathbf{x} \in \mathcal{X}^{seen}, y_{\mathbf{x}} \in \mathcal{Y}^{seen}, \mathbf{a}_{y} \in \mathcal{A}^{seen}\},$$2 where $$\mathcal{D}^{seen}=\{(\mathbf{x}, y_{\mathbf{x}}, \mathbf{a}_{y}) \mid \mathbf{x} \in \mathcal{X}^{seen}, y_{\mathbf{x}} \in \mathcal{Y}^{seen}, \mathbf{a}_{y} \in \mathcal{A}^{seen}\},$$3 is sampled from a $$\mathcal{D}^{seen}=\{(\mathbf{x}, y_{\mathbf{x}}, \mathbf{a}_{y}) \mid \mathbf{x} \in \mathcal{X}^{seen}, y_{\mathbf{x}} \in \mathcal{Y}^{seen}, \mathbf{a}_{y} \in \mathcal{A}^{seen}\},$$4 distribution. The resulting mixed feature is assigned the unknown label. Semantic Based Unknown Generation (SBUG) is more ambitious: it learns a transformed semantic space with a variationally conditioned Wasserstein GAN, samples semantic vectors

$$\mathcal{D}^{seen}=\{(\mathbf{x}, y_{\mathbf{x}}, \mathbf{a}_{y}) \mid \mathbf{x} \in \mathcal{X}^{seen}, y_{\mathbf{x}} \in \mathcal{Y}^{seen}, \mathbf{a}_{y} \in \mathcal{A}^{seen}\},$$5

and optimizes

$$\mathcal{D}^{seen}=\{(\mathbf{x}, y_{\mathbf{x}}, \mathbf{a}_{y}) \mid \mathbf{x} \in \mathcal{X}^{seen}, y_{\mathbf{x}} \in \mathcal{Y}^{seen}, \mathbf{a}_{y} \in \mathcal{A}^{seen}\},$$6

with

$$\mathcal{D}^{seen}=\{(\mathbf{x}, y_{\mathbf{x}}, \mathbf{a}_{y}) \mid \mathbf{x} \in \mathcal{X}^{seen}, y_{\mathbf{x}} \in \mathcal{Y}^{seen}, \mathbf{a}_{y} \in \mathcal{A}^{seen}\},$$7

plus the gradient penalty

$$\mathcal{D}^{seen}=\{(\mathbf{x}, y_{\mathbf{x}}, \mathbf{a}_{y}) \mid \mathbf{x} \in \mathcal{X}^{seen}, y_{\mathbf{x}} \in \mathcal{Y}^{seen}, \mathbf{a}_{y} \in \mathcal{A}^{seen}\},$$8

and the classification loss

$$\mathcal{D}^{seen}=\{(\mathbf{x}, y_{\mathbf{x}}, \mathbf{a}_{y}) \mid \mathbf{x} \in \mathcal{X}^{seen}, y_{\mathbf{x}} \in \mathcal{Y}^{seen}, \mathbf{a}_{y} \in \mathcal{A}^{seen}\},$$9

Unknown semantic embeddings are then synthesized by interpolation in complementary regions of semantic space,

$$\mathcal{D}^{unseen}=\{(\tilde{y}_i, \mathbf{a}_{\tilde{y}_i}) \mid \tilde{y}_i \in \mathcal{Y}^{unseen}, \mathbf{a}_{\tilde{y}_i} \in \mathcal{A}^{unseen}\}.$$0

and used to generate unknown visual features (Marmoreo et al., 2021).

This line of work established two empirical points. First, simply combining a strong GZSL generator with OpenMax does not solve OZSL. Second, unknown generation is beneficial: DUG improved $$\mathcal{D}^{unseen}=\{(\tilde{y}_i, \mathbf{a}_{\tilde{y}_i}) \mid \tilde{y}_i \in \mathcal{Y}^{unseen}, \mathbf{a}_{\tilde{y}_i} \in \mathcal{A}^{unseen}\}.$$1 across CLSWGAN, tf-VAEGAN, and CEZSL, and SBUG often improved over the base VAcWGAN setup (Marmoreo et al., 2021).

ASE (“Adversarial Semantic Embeddings”) reframes the same core problem as ZS-OSR and argues that naïvely chaining a generative ZSL model with an OSR detector fails because synthesized unseen features are generated without any awareness of unknown classes (Li et al., 2023). ASE therefore learns proxy semantic embeddings for unknown classes. Its objective is

$$\mathcal{D}^{unseen}=\{(\tilde{y}_i, \mathbf{a}_{\tilde{y}_i}) \mid \tilde{y}_i \in \mathcal{Y}^{unseen}, \mathbf{a}_{\tilde{y}_i} \in \mathcal{A}^{unseen}\}.$$2

with semantic-space proximity

$$\mathcal{D}^{unseen}=\{(\tilde{y}_i, \mathbf{a}_{\tilde{y}_i}) \mid \tilde{y}_i \in \mathcal{Y}^{unseen}, \mathbf{a}_{\tilde{y}_i} \in \mathcal{A}^{unseen}\}.$$3

and feature-space adversarial loss

$$\mathcal{D}^{unseen}=\{(\tilde{y}_i, \mathbf{a}_{\tilde{y}_i}) \mid \tilde{y}_i \in \mathcal{Y}^{unseen}, \mathbf{a}_{\tilde{y}_i} \in \mathcal{A}^{unseen}\}.$$4

The learned unknown semantic embeddings are passed through a pretrained generative ZSL model to synthesize unknown features, which are then pooled into a single reject label and used to train a $$\mathcal{D}^{unseen}=\{(\tilde{y}_i, \mathbf{a}_{\tilde{y}_i}) \mid \tilde{y}_i \in \mathcal{Y}^{unseen}, \mathbf{a}_{\tilde{y}_i} \in \mathcal{A}^{unseen}\}.$$5-class unknowns-informed classifier (Li et al., 2023).

ASE provides the strongest direct OZSL evidence in the surveyed material. Reported AUROC values are 80.20 on CUB, 81.99 on AWA2, 72.78 on FLO, and 72.69 on SUN, and the paper states a 12.42% reduction in FPR95 on AWA2 relative to the best baseline. It also extends to generalized ZS-OSR and ZS-OOD, where unknown classes come from different datasets (Li et al., 2023).

5. Open-world label spaces, compositionality, and feasibility modeling

A major branch of OZSL-related research studies openness as expansion of the candidate label space rather than only as explicit unknown rejection. In OW-CZSL, the label set is the full Cartesian product

$$\mathcal{D}^{unseen}=\{(\tilde{y}_i, \mathbf{a}_{\tilde{y}_i}) \mid \tilde{y}_i \in \mathcal{Y}^{unseen}, \mathbf{a}_{\tilde{y}_i} \in \mathcal{A}^{unseen}\}.$$6

and open world is defined by

$$\mathcal{D}^{unseen}=\{(\tilde{y}_i, \mathbf{a}_{\tilde{y}_i}) \mid \tilde{y}_i \in \mathcal{Y}^{unseen}, \mathbf{a}_{\tilde{y}_i} \in \mathcal{A}^{unseen}\}.$$7

CompCos scores images against compositions by cosine similarity,

$$\mathcal{D}^{unseen}=\{(\tilde{y}_i, \mathbf{a}_{\tilde{y}_i}) \mid \tilde{y}_i \in \mathcal{Y}^{unseen}, \mathbf{a}_{\tilde{y}_i} \in \mathcal{A}^{unseen}\}.$$8

and supplements this with feasibility scores

$$\mathcal{D}^{unseen}=\{(\tilde{y}_i, \mathbf{a}_{\tilde{y}_i}) \mid \tilde{y}_i \in \mathcal{Y}^{unseen}, \mathbf{a}_{\tilde{y}_i} \in \mathcal{A}^{unseen}\}.$$9

combined as

$\mathcal{X}^{unseen} \cup \mathcal{X}^{unknown},$0

The paper reports that the average mixing function works best, and uses feasibility both for hard masking,

$\mathcal{X}^{unseen} \cup \mathcal{X}^{unknown},$1

and for feasibility-aware training (Mancini et al., 2021).

The empirical conclusion of OW-CZSL is severe: all methods deteriorate sharply when moving from closed-world to open-world evaluation. On MIT States, LE+ unseen accuracy drops from $\mathcal{X}^{unseen} \cup \mathcal{X}^{unknown},$2 to $\mathcal{X}^{unseen} \cup \mathcal{X}^{unknown},$3, SymNet drops from $\mathcal{X}^{unseen} \cup \mathcal{X}^{unknown},$4 to $\mathcal{X}^{unseen} \cup \mathcal{X}^{unknown},$5, and TMN drops from $\mathcal{X}^{unseen} \cup \mathcal{X}^{unknown},$6 to $\mathcal{X}^{unseen} \cup \mathcal{X}^{unknown},$7. CompCos recovers part of this loss, improving over SymNet by $\mathcal{X}^{unseen} \cup \mathcal{X}^{unknown},$8 unseen accuracy and doubling $\mathcal{X}^{unseen} \cup \mathcal{X}^{unknown},$9 from $$\phi^{open} : \mathbf{x} \mapsto \mathcal{O}, \qquad \mathcal{O} = \mathcal{Y}^{unseen} \cup \{\text{'unknown'}\}.$$0 to $$\phi^{open} : \mathbf{x} \mapsto \mathcal{O}, \qquad \mathcal{O} = \mathcal{Y}^{unseen} \cup \{\text{'unknown'}\}.$$1 on MIT States (Mancini et al., 2021).

Feasibility with LLM (FLM) replaces static feasibility priors with LLM-based semantic judgments. For a state-object pair $$\phi^{open} : \mathbf{x} \mapsto \mathcal{O}, \qquad \mathcal{O} = \mathcal{Y}^{unseen} \cup \{\text{'unknown'}\}.$$2, the feasibility score is

$$\phi^{open} : \mathbf{x} \mapsto \mathcal{O}, \qquad \mathcal{O} = \mathcal{Y}^{unseen} \cup \{\text{'unknown'}\}.$$3

or, with in-context guidance,

$$\phi^{open} : \mathbf{x} \mapsto \mathcal{O}, \qquad \mathcal{O} = \mathcal{Y}^{unseen} \cup \{\text{'unknown'}\}.$$4

where

$$\phi^{open} : \mathbf{x} \mapsto \mathcal{O}, \qquad \mathcal{O} = \mathcal{Y}^{unseen} \cup \{\text{'unknown'}\}.$$5

Thresholding then produces the candidate set

$$\phi^{open} : \mathbf{x} \mapsto \mathcal{O}, \qquad \mathcal{O} = \mathcal{Y}^{unseen} \cup \{\text{'unknown'}\}.$$6

This feasibility filter consistently improves CLIP, CoOp, and CSP across MIT-States, UT-Zappos, and C-GQA, and the paper reports especially strong gains on UT-Zappos (Kim et al., 16 May 2025).

The 2024 unified OW-CZSL framework pushes this line further with stronger multimodal interaction. It concatenates image patch embeddings and text embeddings as transformer input, introduces a TopK Embedding Selection Module to retain image-relevant textual candidates, and uses a Sparse Linear Compositor to produce attribute, object, and final composition predictions. It also introduces a hybrid procedure that combines joint composition learning with independent primitive learning, and the abstract reports state-of-the-art performance in OW-CZSL on three datasets while surpassing Large Vision LLMs in two datasets (Jayasekara et al., 2024).

A broader, non-compositional precursor is vocabulary-informed learning, which trains a single semantic-prototype model using both labeled source classes and a much larger unlabeled semantic vocabulary. It defines inference over supervised, zero-shot, generalized zero-shot, and open-set regimes by changing only the candidate label set,

$$\phi^{open} : \mathbf{x} \mapsto \mathcal{O}, \qquad \mathcal{O} = \mathcal{Y}^{unseen} \cup \{\text{'unknown'}\}.$$7

with open set given by

$$\phi^{open} : \mathbf{x} \mapsto \mathcal{O}, \qquad \mathcal{O} = \mathcal{Y}^{unseen} \cup \{\text{'unknown'}\}.$$8

Its weighted maximum-margin objective incorporates constraints against both seen and unsupervised vocabulary atoms, and the experiments scale to an open vocabulary of about 310K words/phrases (Fu et al., 2023). The method is highly relevant to open-vocabulary OZSL, but it has no explicit unknown rejection rule and always assigns the nearest prototype.

6. Semantic backbones, empirical trends, and limitations

A recurring theme in OZSL is that open-world handling is only as good as the semantic backbone it rests on. OntoZSL replaces limited priors such as plain attributes or word embeddings with ontology-based semantics represented as a multi-relational RDF graph

$$\phi^{open} : \mathbf{x} \mapsto \mathcal{O}, \qquad \mathcal{O} = \mathcal{Y}^{unseen} \cup \{\text{'unknown'}\}.$$9

and learns text-aware ontology embeddings with translational scores such as

$\mathcal{S},\qquad \mathcal{U},\qquad \Omega,$0

These embeddings are then fed into a conditional GAN,

$\mathcal{S},\qquad \mathcal{U},\qquad \Omega,$1

to synthesize unseen features (Geng et al., 2021). OntoZSL is not an open-set or open-world method in the strict sense, but it is directly relevant as a richer semantic backbone. The paper states that ontology-based class semantics outperform previous priors such as word embeddings by an average of 12.4 accuracy points in standard ZSL across two example datasets.

On the visual side, ZeroMamba shows that stronger visual-semantic transfer can materially improve conventional ZSL and GZSL even without solving true OZSL. Its core components are Semantic-aware Local Projection,

$\mathcal{S},\qquad \mathcal{U},\qquad \Omega,$2

Global Representation Learning,

$\mathcal{S},\qquad \mathcal{U},\qquad \Omega,$3

and Semantic Fusion,

$\mathcal{S},\qquad \mathcal{U},\qquad \Omega,$4

trained with

$\mathcal{S},\qquad \mathcal{U},\qquad \Omega,$5

Its generalized prediction rule includes an unseen-class calibration term,

$\mathcal{S},\qquad \mathcal{U},\qquad \Omega,$6

which is OZSL-relevant because seen-class bias is a prerequisite problem for any open-world extension (Hou et al., 2024).

Older ZSL formulations remain methodologically important. The “missing data problem” view models each unseen class as a Gaussian in feature space,

$\mathcal{S},\qquad \mathcal{U},\qquad \Omega,$7

initialized from semantically transferred virtual signatures and refined by EM (Zhao et al., 2016). SynC and EXEM synthesize unseen classifiers

$\mathcal{S},\qquad \mathcal{U},\qquad \Omega,$8

or predicted exemplars

$\mathcal{S},\qquad \mathcal{U},\qquad \Omega,$9

for predefined unseen classes (Changpinyo et al., 2018). These methods are foundational because OZSL still needs strong candidate unseen-class models, but they remain closed-world.

Taken together, the literature indicates several stable conclusions. Closed-world assumptions can dramatically overestimate performance: open-world compositional evaluation causes severe degradation (Mancini et al., 2021). Naïve combinations of ZSL and OSR perform badly in distinguishing unseen-class and unknown-class samples (Li et al., 2023). Semantic quality is a dominant variable: ontology-grounded semantics, exemplar prediction, and vocabulary-informed margins all improve transfer to predefined unseen classes (Geng et al., 2021, Changpinyo et al., 2018, Fu et al., 2023). Feasibility modeling helps when the open label space is large and noisy, whether through primitive similarities, LLM priors, or multimodal TopK filtering (Mancini et al., 2021, Kim et al., 16 May 2025, Jayasekara et al., 2024). At the same time, many ostensibly related methods stop short of OZSL because they retain a closed candidate unseen set, provide no reject option, or collapse all novelty into nearest-prototype assignment (Geng et al., 2021, Hou et al., 2024, Fu et al., 2023).

The main unresolved issue is therefore not merely zero-shot transfer, but open-world calibration under incomplete semantic support. The surveyed works suggest several non-exclusive routes: synthesize proxy unknowns, model feasibility before visual ranking, improve semantic structure with ontologies or large vocabularies, and strengthen visual-semantic alignment so that rejection scores are less confounded by poor transfer. This suggests that OZSL is not a single method family but a junction of ZSL, OSR, open-vocabulary recognition, and structured semantic reasoning (Marmoreo et al., 2021).