Zero-Shot Prompting
Large LLMs (language models) such as GPT-5, Claude 3, or Qwen are currently designed to follow instructions and have been trained on large datasets. This large-scale training enables these models to handle certain tasks in a “zero-shot” manner. Zero-shot prompting refers to a prompt that contains no examples or demonstrations and directly instructs the model to perform a task without additional examples.
Below is an example of zero-shot prompting (text classification):
Klassifizieren Sie den Text als neutral, negativ oder positiv.
Text: Ich finde den Urlaub okay.
Empfindung:NeutralNote that the prompt above contains no example texts together with their classifications; the LLM already interprets the term “Empfindung,” illustrating zero-shot capabilities in practice.
Instruction tuning has been shown to improve zero-shot learning Wei et al. (2022)1. Instruction tuning essentially refers to fine-tuning models on datasets described by instructions. In addition, RLHF (Reinforcement Learning from Human Feedback) has been adapted according to Christiano et al. (2023)2 to scale instruction tuning by aligning the model more closely with human preferences. This recent development forms the basis for models such as ChatGPT. The different approaches and methods are explained in the following sections.
If zero-shot prompting proves insufficient, it is recommended to provide demonstrations or examples in the prompt, which leads to few-shot prompting. The next section introduces few-shot prompting.
Wei, J., Bosma, M., Zhao, V. Y., Guu, K., Yu, A. W., Lester, B., Du, N., Dai, A. M., & Le, Q. V. (2022). Finetuned Language Models Are Zero-Shot Learners (No. arXiv:2109.01652). arXiv. https://doi.org/10.48550/arXiv.2109.01652 ↩︎
Christiano, P., Leike, J., Brown, T. B., Martic, M., Legg, S., & Amodei, D. (2023). Deep reinforcement learning from human preferences (No. arXiv:1706.03741). arXiv. https://doi.org/10.48550/arXiv.1706.03741 ↩︎