Few-Shot Zero-Glance: Hybrid Learning Paradigm

• Few-Shot Zero-Glance is a learning paradigm that fuses zero-shot semantic priors with few-shot prototyping to rapidly adapt to new classes with minimal labeled data.
• It leverages pre-trained language and multi-modal models to construct class prototypes via prompt-based fusion and simple aggregation methods.
• The framework effectively applies across various domains—including vision, audio, and graphs—demonstrating strong performance with minimal supervision.

Few-Shot Zero-Glance denotes a class of methodologies characterized by hybridization of zero-shot semantic priors and few-shot prototyping within a unified episodic meta-learning framework. These methods aim to maximize data efficiency, leveraging the generalization potential of pre-trained language or multi-modal models alongside rapid adaptation from a handful of labeled instances. “Zero-glance” signifies a setting where the model operates either without exposure to true labeled examples (pure zero-shot) or with extremely sparse supervision—facilitating prompt-based or prototype-based inference as the learning protocol transitions from zero- to few-shot regimes (Zhou et al., 10 Jan 2024). The approach has proliferated across visual, textual, graph, and audio domains, with distinct instantiations in LLMs, vision-LLMs, open-vocabulary detectors, foundation models for anomaly detection, and generative ZSL frameworks.

1. Core Principles and Definitions

Few-Shot Zero-Glance methods fundamentally exploit the compositionality of semantic embeddings derived from large pre-trained models. When $K=0$ (zero-shot), prediction is based solely on a semantic prior—typically a prompt-encoded or attribute-driven vector derived from a LLM or supervised ontology. When $K>0$ (few-shot), prototype construction or parameter adaptation incorporates the limited available data, fusing visual and semantic features—usually via simple addition, averaging, or cross-modal attention.

The “zero-glance” notion emphasizes that the model has never seen actual ground-truth examples for the target (novel) classes at training time, or operates with extremely constrained labeled support sets at test time (often $K\ll 10$). Prominent frameworks explicitly harness the zero-shot capacity of frozen pre-trained LMs with learnable prompts, and further aggregate visual–textual feature representations by simple addition, obviating complex fusion modules (Zhou et al., 10 Jan 2024).

2. Representative Methodologies

2.1 Prompt-Conditioned Prototypical Fusion

A prototypical Few-Shot Zero-Glance pipeline (e.g., “SimpleFSL”) involves:

• Prompt-based zero-shot prior: Predefined or learnable prompts $P_i$ representing class $c_i$, passed through a frozen LLM $g(\cdot)$ to yield embeddings $t_i = g(P_i)$.
• Prompt–visual adaptation: Each class embedding $t_i$ is mapped (with, e.g., an MLP) to the visual feature space for compatibility.
• Prototype construction: For class $c_i$ with $K$ support samples, the class prototype is:

$$p_i = \frac{1}{K} \sum_{j=1}^{K} \big(f_\Theta(x_{i,j}) + z_i\big)$$

with $z_i$ the adapted prompt embedding and $f_\Theta(x)$ the image embedding. For pure zero-shot ($K=0$), $p_i = z_i$.

• Non-parametric classification: Query-image features are compared via cosine similarity with prototypes. The predicted label is $$y^* = \arg\max_i \langle f_\Theta(x_q), p_i \rangle / \tau$$, where $\tau$ is a temperature parameter.
• Self-ensemble/distillation: Blending the prompt-fused logits with visual-only prototypes (and distilling via symmetric KL divergence) yields additional accuracy gains (Zhou et al., 10 Jan 2024).

2.2 Augmented In-Context Demonstrations

For LLMs, Z-ICL synthesizes in-context “pseudo-demonstrations” by retrieving unlabeled corpus neighbors, assigning random labels (with label synonyms to avoid leakage), and presenting these as artificial prompt sequences. Models thus “glance” at plausible demonstrations constructed without access to true task labels or examples, closing the gap to few-shot ICL (Lyu et al., 2022).

2.3 Prototype Averaging and Adaptive Aggregation

Across modalities (audio, image, 3D), learned class prototypes in embedding space constructed from few-shot samples consistently outperform text/prompt-based baselines. The fusion is typically an (optionally normalized) mean of instance embeddings; performance saturates or slightly decreases beyond a moderate $K$, with just a handful of examples (e.g., $K=5$–$10$) sufficing to surpass zero-shot performance (Taylor et al., 26 Jul 2025, Lin et al., 30 Apr 2024).

3. Applications Across Modalities

The Few-Shot Zero-Glance paradigm is agnostic to input and output modality, provided cross-modal embeddings and/or semantic prompt alignment can be established. Representative domains include:

• Visual classification: Prompt-fused prototype methods and few-shot prompt tuning outperform purely supervised FSL and prompt-tuning, with minimal adaptation (Zhou et al., 10 Jan 2024).
• Audio classification: Prototype averaging in embedding space leverages a few real-class examples, replacing fragile textual prompts and yielding improved mean average precision and accuracy metrics over large audio benchmarks (Taylor et al., 26 Jul 2025).
• Object detection: Open-vocabulary detectors converted to few-shot instance detectors simply by replacing text-derived embeddings with image-derived prototypes; no fine-tuning required (Crulis et al., 21 Oct 2024).
• Foundation graph anomaly detection: Residual-based prototype matching yields both zero-shot and lightweight few-shot anomaly detection that generalizes strongly across graphs (Qiao et al., 13 Feb 2025).
• Action recognition and multi-view 3D shape classification: Temporal attention and prompt-enhanced aggregation provide state-of-the-art performance by combining semantic guidance and data-driven improvements accessible even with few labeled samples (Bishay et al., 2019, Lin et al., 30 Apr 2024).

4. Experimental Findings and Comparative Analyses

Empirical studies consistently demonstrate tangible and sometimes substantial gains versus standard zero-shot or supervised few-shot approaches upon integrating Few-Shot Zero-Glance strategies. Notable results include:

Method	1-shot miniImageNet	5-shot miniImageNet	1-shot CIFAR-FS	5-shot CIFAR-FS
ProtoNet	60.3%	80.5%	72.2%	83.5%
SP-CLIP	72.4%	83.2%	82.2%	88.2%
SimpleFSL	74.8%	83.3%	84.8%	88.8%
SimpleFSL++	75.6%	83.9%	85.1%	89.1%

(Zhou et al., 10 Jan 2024)

On ESC-50 audio, AVG prototypes match or surpass zero-shot CLAP23 accuracy with as few as 5–6 examples/class (zero-shot 0.948, FS_AVG(K=10) 0.970) (Taylor et al., 26 Jul 2025). In 3D shape recognition, prompt-enhanced view aggregation enables CLIP-based zero-shot accuracy to rise significantly with only a few examples, aided further by feature distillation (Lin et al., 30 Apr 2024).

5. Design Considerations and Broader Implications

Feature fusion simplicity: Direct addition of visual and textual features is highly effective, often outperforming elaborate fusion strategies, particularly when the LM backbone is reasonably aligned via learned prompts or projection (Zhou et al., 10 Jan 2024).

Prompt design and learnability: Learnable context tokens outperform fixed or manually-designed prompts. Proper prompt adaptation is essential for transferring zero-shot priors effectively to the few-shot regime (Zhou et al., 10 Jan 2024).

Prototype aggregation details: Both uniform and normalized averaging are robust; normalized averaging may better match cosine-based contrastive pre-training objectives (Taylor et al., 26 Jul 2025).

Regularization and calibration: Cosine similarity is critical in embedding-based classification. Mutual-information-based dimensionality reduction can aid LDA-based few-shot classifiers, though improvements over simple averaging are modest (Taylor et al., 26 Jul 2025).

Scalability and extensibility: Prototype-based classifiers are incremental—extending to new classes or updating prototypes as data accumulates does not require backbone retraining. The paradigm generalizes to multi-label and hierarchical tasks, few-shot prompt tuning, and adapts cleanly to large-scale graphs and multi-modal data (Qiao et al., 13 Feb 2025, Crulis et al., 21 Oct 2024).

6. Limitations and Open Problems

Few-Shot Zero-Glance approaches are not universally optimal. Existing frameworks can be sensitive to the semantics and informativeness of textual/semantic priors; prototype-based methods may still underperform when fine-grained alignment is absent or when prompt/attribute sets poorly cover target domains (Taylor et al., 26 Jul 2025, Zhou et al., 10 Jan 2024). Extreme domain shifts or label noise can also compromise both zero-shot and few-shot accuracy. Automated prompt or synonym selection and the extension beyond classification (e.g., to sequence labeling or generation) remain significant challenges (Lyu et al., 2022).

7. Outlook and Future Directions

Current research suggests several promising directions:

• Automated prompt/synonym mining: To reduce human curation and further close the zero/few-shot performance gap (Lyu et al., 2022).
• Multi-modal and hierarchical extensions: Leveraging semantic prototypes for complex or structured outputs, including multi-label, hierarchical, or anomaly detection in graphs and temporal data (Qiao et al., 13 Feb 2025).
• Learnable fusion strategies: Composing multiple prompt-conditioned prototypes or integrating side information in a data-driven manner.
• Robustness enhancement: Addressing domain shift, heterogeneity, and adversarial effects, particularly when class priors are weak or insufficient (Zhou et al., 10 Jan 2024).
• Datasets and benchmarks: Expansion to broader modalities and task types will clarify the generality and boundaries of Few-Shot Zero-Glance paradigms.

In summary, Few-Shot Zero-Glance represents a unifying and highly extensible framework for minimal-supervision learning, capitalizing on the complementarity between frozen semantic priors and rapid adaptation from few-shot instances. Its simplicity, empirical effectiveness, and conceptual alignment with modern pre-trained models underlie its continued adoption across machine learning domains (Zhou et al., 10 Jan 2024, Taylor et al., 26 Jul 2025, Lyu et al., 2022, Qiao et al., 13 Feb 2025, Crulis et al., 21 Oct 2024).