Few-Shot and Zero-Shot Learning Cheat Sheet

Updated 2026-05-25

Next Topic: Foundation Models in AI Cheat Sheet

🧠Study flashcards on this topic115 cards · spaced repetition→

Few-shot and zero-shot learning are machine learning paradigms that enable models to generalize to new tasks or classes with minimal labeled examples—ranging from none (zero-shot) to a small handful (few-shot). These approaches are foundational to in-context learning in large language models and meta-learning in computer vision, where models learn to adapt quickly by transferring knowledge from prior experience rather than requiring extensive task-specific training. The key challenge is to design representations, prompting strategies, and meta-learning algorithms that maximize generalization from extremely limited supervision, making these techniques essential for real-world applications where labeled data is scarce, expensive, or rapidly changing. Understanding the nuances between demonstration selection, calibration methods, and architectural choices directly impacts whether a model performs near state-of-the-art or random-guess accuracy on new tasks.

What This Cheat Sheet Covers

This topic spans 15 focused tables and 117 indexed concepts, 115 flashcards. Below is a complete table-by-table outline of this topic, spanning foundational concepts through advanced details.

Table 1: Core Learning ParadigmsTable 2: Prompting Strategies for LLMsTable 3: Meta-Learning ArchitecturesTable 4: Demonstration Selection StrategiesTable 5: Task Formulation & TrainingTable 6: Inference ApproachesTable 7: Calibration & Mitigation MethodsTable 8: Prompt Design ComponentsTable 9: Advanced Prompt TechniquesTable 10: Cross-Domain & Specialized FSLTable 11: Common Challenges & Failure ModesTable 12: Vision BenchmarksTable 13: NLP BenchmarksTable 14: Practical Considerations for DeploymentTable 15: Real-World Applications

Quick IndexSubscribe to unlock

A jump-to index of every table row in this cheat sheet.

Mind MapSubscribe to unlock

An interactive map of every table and concept in this topic.

Table 1: Core Learning Paradigms

The foundational vocabulary of few-shot and zero-shot learning defines the problem structure, data splits, and the main training philosophies. Knowing how these paradigms relate helps you choose the right approach — prompting a large LLM vs. training a metric-based model vs. full meta-learning — for your data and compute constraints.

Paradigm	Example	Description
Zero-shot learning	`Classify sentiment: "I loved it!" → Positive`	• Model performs a task with no examples, relying entirely on pre-trained knowledge and task instructions • works best for generalized tasks that match training distribution.
One-shot learning	`Example: "Great movie" → Positive` `Classify: "Terrible film"`	• Model receives exactly one example per class before inference • bridges zero-shot and few-shot by providing minimal demonstration of desired behavior.
Few-shot learning	`3 examples:` `"Loved it" → Pos` `"Hated it" → Neg` `"Okay" → Neu`	• Model learns from 2–10 labeled examples per class • significantly improves performance over zero-shot for specialized or nuanced tasks.
K-shot learning	`k=5: 5 examples per class`	• Formalization of few-shot where k specifies the exact number of examples per class • typically k ∈ {1, 2, 5, 10} in research benchmarks.
N-way K-shot	`5-way 1-shot: 5 classes, 1 example each`	• Standard task formulation where model discriminates between N classes given K examples per class • defines the few-shot classification problem structure.
In-context learning (ICL)	`Prompt with examples inline, no weight updates`	• LLM capability to learn from demonstrations in the prompt at inference time without gradient updates • encompasses zero-shot, one-shot, and few-shot prompting.
Generalized Zero-Shot Learning (GZSL)	`Test on both seen and unseen classes simultaneously`	• More realistic variant of ZSL where the model must classify both seen and unseen classes at test time • evaluated using harmonic mean H of seen and unseen accuracy to prevent bias toward seen classes.
Meta-learning	`MAML learns initialization for fast adaptation`	Training paradigm that learns to learn—optimizes for rapid adaptation to new tasks from few examples by learning good parameter initializations or similarity metrics.
Episodic training	`Each episode: sample support set + query set`	• Training strategy where model sees many synthetic few-shot tasks (episodes) to learn how to generalize from limited data • mimics test-time conditions.

Table 2: Prompting Strategies for LLMs

These prompting techniques control how demonstrations and instructions are presented to a language model at inference time—no weight updates involved. Choosing between zero-shot, few-shot, and many-shot depends on task complexity, context window size, and available examples; picking the right strategy often matters more than model size.

Technique	Example	Description
Zero-shot prompting	`"Translate to French: Hello"`	• Direct instruction with no demonstrations • model relies on pre-training • sensitive to prompt phrasing; benefits from clear task framing.

The rest of this section is part of the full cheat sheet

Table 3: Meta-Learning Architectures

Meta-learning algorithms learn how to learn—they train across many tasks so the model can adapt quickly at test time using only a few labeled examples. The choice between metric-based methods (fast inference, no updates) and optimization-based methods (slower, gradient updates at test time) is the central design decision.

Method	Example	Description
Model-Agnostic Meta-Learning (MAML)	`θ′ = θ − α∇L(θ, task_i)`	• Learns an initialization that enables fast adaptation with few gradient steps • optimizes for good performance across tasks after minimal fine-tuning.

The rest of this section is part of the full cheat sheet

Table 4: Demonstration Selection Strategies

How you choose which examples to include in a few-shot prompt has as much impact as the model itself. Poor selection can tank performance; smart, diverse, or query-matched selection can push few-shot results close to full fine-tuning.

Strategy	Example	Description
Semantic similarity	`Use BERT similarity to select closest examples`	• Retrieves examples most similar to the query using embedding-based distance (e.g., cosine similarity) • improves task-specific relevance.

The rest of this section is part of the full cheat sheet

Table 1: Core Learning Paradigms

Paradigm	Example	Description
Zero-shot learning	`Classify sentiment: "I loved it!" → Positive`	• Model performs a task with no examples, relying entirely on pre-trained knowledge and task instructions • works best for generalized tasks that match training distribution.
One-shot learning	`Example: "Great movie" → Positive` `Classify: "Terrible film"`	• Model receives exactly one example per class before inference • bridges zero-shot and few-shot by providing minimal demonstration of desired behavior.
Few-shot learning	`3 examples:` `"Loved it" → Pos` `"Hated it" → Neg` `"Okay" → Neu`	• Model learns from 2–10 labeled examples per class • significantly improves performance over zero-shot for specialized or nuanced tasks.
K-shot learning	`k=5: 5 examples per class`	• Formalization of few-shot where k specifies the exact number of examples per class • typically k ∈ {1, 2, 5, 10} in research benchmarks.
N-way K-shot	`5-way 1-shot: 5 classes, 1 example each`	• Standard task formulation where model discriminates between N classes given K examples per class • defines the few-shot classification problem structure.
In-context learning (ICL)	`Prompt with examples inline, no weight updates`	• LLM capability to learn from demonstrations in the prompt at inference time without gradient updates • encompasses zero-shot, one-shot, and few-shot prompting.
Generalized Zero-Shot Learning (GZSL)	`Test on both seen and unseen classes simultaneously`	• More realistic variant of ZSL where the model must classify both seen and unseen classes at test time • evaluated using harmonic mean H of seen and unseen accuracy to prevent bias toward seen classes.
Meta-learning	`MAML learns initialization for fast adaptation`	Training paradigm that learns to learn—optimizes for rapid adaptation to new tasks from few examples by learning good parameter initializations or similarity metrics.
Episodic training	`Each episode: sample support set + query set`	• Training strategy where model sees many synthetic few-shot tasks (episodes) to learn how to generalize from limited data • mimics test-time conditions.

Table 2: Prompting Strategies for LLMs

Technique	Example	Description
Zero-shot prompting	`"Translate to French: Hello"`	• Direct instruction with no demonstrations • model relies on pre-training • sensitive to prompt phrasing; benefits from clear task framing.

Table 3: Meta-Learning Architectures

Method	Example	Description
Model-Agnostic Meta-Learning (MAML)	`θ′ = θ − α∇L(θ, task_i)`	• Learns an initialization that enables fast adaptation with few gradient steps • optimizes for good performance across tasks after minimal fine-tuning.

Table 4: Demonstration Selection Strategies

Strategy	Example	Description
Semantic similarity	`Use BERT similarity to select closest examples`	• Retrieves examples most similar to the query using embedding-based distance (e.g., cosine similarity) • improves task-specific relevance.