Example-driven prompts leverage demonstrations to guide large language models (LLMs) in producing precise and contextually relevant outputs, forming a critical component of prompt engineering. This category includes zero-shot prompting, one-shot prompting, and few-shot prompting, each offering varying degrees of guidance for research tasks. These techniques enable researchers to tailor LLM outputs without extensive model training, balancing flexibility and specificity.
Zero-shot prompting enables LLMs to perform tasks without task-specific examples, relying solely on a descriptive prompt and the model’s pre-trained knowledge (Brown et al. 2020). This approach is grounded in the model’s ability to generalise from its training corpus, applying learned patterns to novel tasks. In zero-shot prompting, the model infers the desired output from the prompt’s instructions. For example, a researcher might prompt an LLM to “Extract the primary diagnosis from this clinical note: ‘Patient presents with persistent cough and fever.’” The model, leveraging its pre-trained medical knowledge, identifies the diagnosis (e.g., “Possible pneumonia”) without prior examples (Sivarajkumar et al. 2024). This simplicity makes zero-shot prompting ideal for rapid experimentation, particularly in domains with scarce labelled data, such as rare disease analysis. Zero-shot prompting is straightforward and requires no additional data, making it suitable for low-resource settings (Li 2023). However, its performance varies with task complexity and the model’s pre-training exposure. Nuanced tasks, such as extracting ambiguous clinical entities, often yield suboptimal results (Sivarajkumar et al. 2024). Techniques like instruction tuning, where models are fine-tuned on diverse task instructions, and reinforcement learning with human feedback (RLHF) can enhance zero-shot performance by improving generalisation and response quality (Chen et al. 2025).
One-shot prompting provides a single example alongside the instruction, offering a reference for the task and improving the model’s ability to infer requirements (Wei et al., 2022). One-shot prompting leverages in-context learning, where the model uses the example to align its output with the desired format. For instance, in named entity recognition (NER), a prompt might include: “Text: ‘The protein Mitochondrial Enzyme X (MEX) regulates metabolism.’ Entity: Mitochondrial Enzyme X (MEX). Now extract the entity from: ‘Type 2 Diabetes Mellitus (T2DM) is a chronic condition.’” The model infers “Type 2 Diabetes Mellitus (T2DM)” as the entity, guided by the example (Cheng et al. 2024). One-shot prompting balances simplicity and improved performance, making it effective for tasks with limited data, such as analysing specialised scientific texts (Ha et al. 2025). However, its success hinges on the example’s quality; an unrepresentative example can mislead the model (Reynolds & McDonell 2021). Additionally, one-shot prompting may struggle with tasks requiring complex reasoning or multiple output formats.
Few-shot prompting provides multiple examples within the prompt, enabling the model to discern patterns and produce more accurate outputs (Brown et al. 2020). Few-shot prompting enhances in-context learning by offering a richer context. For a natural language inference (NLI) task, a prompt might include: “Premise: ‘The drug reduces inflammation.’ Hypothesis: ‘The drug alleviates pain.’ Relationship: Neutral. Premise: ‘The patient received chemotherapy.’ Hypothesis: ‘The patient underwent surgery.’ Relationship: Neutral. Premise: ‘The gene promotes cell growth.’ Hypothesis: ‘The gene inhibits apoptosis.’ Relationship: Contradiction. Now: Premise: ‘The compound binds to Receptor Z.’ Hypothesis: ‘The compound activates Receptor Z.’ Relationship:” The model infers “Entailment” based on the examples (Schick and Schütze 2022). Few-shot prompting significantly outperforms zero- and one-shot methods for complex tasks, reducing the need for large labelled datasets (Bahrami et al. 2023). However, its effectiveness depends on example quality and diversity, and longer prompts increase computational costs (Sanh et al. 2022). Context window limitations also restrict the number of examples, posing challenges for very large models.
References:
1. Bahrami, Morteza, Muharram Mansoorizadeh, and Hassan Khotanlou. 2023. Few-shot Learning with Prompting Methods. In 2023 6th International Conference on Pattern Recognition and Image Analysis (IPRIA), Qom, Iran, 1–5. IEEE. https://doi.org/10.1109/IPRIA59240.2023.10147172 ^ Back
2. Brown, Tom, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D. Kaplan, Prafulla Dhariwal, Arvind Neelakantan, et al. 2020. ‘Language Models Are Few-Shot Learners’. Advances in Neural Information Processing Systems 33: 1877–1901. ^ Back
3. Chen, Banghao, Zhaofeng Zhang, Nicolas Langrené, and Shengxin Zhu. 2025. Unleashing the Potential of Prompt Engineering for Large Language Models. Patterns 6 (6): 101260. https://doi.org/10.1016/j.patter.2025.101260 ^ Back
4. Cheng, Qi, Liqiong Chen, Zhixing Hu, Juan Tang, Qiang Xu, and Binbin Ning. 2024. A Novel Prompting Method for Few-Shot NER via LLMs. Natural Language Processing Journal 8: 100099. https://doi.org/10.1016/j.nlp.2024.100099 ^ Back
5. Ha, Junwoo, Hyunjun Kim, Sangyoon Yu, Haon Park, Ashkan Yousefpour, Yuna Park, and Suhyun Kim. 2025. One-Shot is Enough: Consolidating Multi-Turn Attacks into Efficient Single-Turn Prompts for LLMs. arXiv preprint arXiv:2503.04856. https://arxiv.org/abs/2503.04856 ^ Back
6. Li, Yinheng. 2023. A Practical Survey on Zero-shot Prompt Design for In-context Learning. In Proceedings of the 14th International Conference Recent Advances in Natural Language Processing (RANLP 2023), 637–643. Varna, Bulgaria. arXiv preprint arXiv:2309.13205. https://doi.org/10.48550/arXiv.2309.13205 ^ Back
7. Reynolds, Laria, and Kyle McDonell. 2021. Prompt Programming for Large Language Models: Beyond the Few-Shot Paradigm. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems, pp. 1–7. https://doi.org/10.1145/3411763.3450381 ^ Back
8. Sanh, Victor, Albert Webson, Colin Raffel, Stephen Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, et al. 2022. Multitask Prompted Training Enables Zero-Shot Task Generalization. In Proceedings of the International Conference on Learning Representations (ICLR), Spotlight paper. https://openreview.net/forum?id=9Vrb9D0WI4 ^ Back
9. Schick, Timo, and Hinrich Schütze. 2022. True Few-Shot Learning with Prompts—A Real-World Perspective. Transactions of the Association for Computational Linguistics 10: 716–731. https://aclanthology.org/2022.tacl-1.38/ ^ Back
10. Sivarajkumar, Sonish, Mark Kelley, Alyssa Samolyk-Mazzanti, Shyam Visweswaran, and Yanshan Wang. 2024. An Empirical Evaluation of Prompting Strategies for Large Language Models in Zero-Shot Clinical Natural Language Processing: Algorithm Development and Validation Study. JMIR Medical Informatics 12: e55318. https://doi.org/10.2196/55318 ^ Back
