Conversational Few-Shot Prompting: Rethinking Few-Shot Prompting for Chat Language Model

Bowen Qin, Duanyu Feng, Xi Yang

Keywords: Few-shot learning; In-context learning; Large language model;

TL;DR: A new few-shot prompting method for chat-based language models.

Abstract:

In-context learning, also referred to as few-shot learning, enables language models to adapt to tasks using a limited number of examples embedded in the prompt. Traditional approaches typically present all examples in a single prompt, which works well for pre-trained base models. However, the application of this method to instruction-tuned chat models, such as ChatGPT, remains underexplored.
In this paper, we introduce a novel conversational few-shot prompting technique, which structures few-shot examples as multi-turn conversation between the user and the assistant, rather than a single input prompt. This conversational framing better aligns with the interactive nature of chat models, enhancing their instruction-following abilities and generalization across tasks.
Through experiments on various benchmarks, we demonstrate that this approach significantly improves performance, particularly in low-shot scenarios, compared to traditional few-shot prompting. Our results suggest that this method provides a more flexible and robust way to leverage few-shot examples in instruction-tuned chat models, improving task performance without the need for additional fine-tuning, reducing prompt sensitivity, and offering potential for diverse applications.

Primary Area: applications to computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5999