Text as parameter: interactive prompt optimisation for large language models
Keywords: Large language model, prompt optimisation, dialogue
Abstract: Large language models (LLMs) can handle a variety of tasks conditioned on natural language instructions. While fine-tuning improves task-specific performance, adjusting the model weights of LLMs requires a huge amount of computational resources, and it is impractical for real-time updates. Alternatively, prompting allows LLMs to adapt to a broad range of tasks without the need for computationally intensive gradient-based optimisation.
However, crafting effective prompts remains a challenge, to the extent that it is even unclear if expert in-domain knowledge is what is needed or experience in writing prompts or something else.
Approaches like meta-prompting and self-feedback seek to alleviate this burden, but they rely primarily on a numerical feedback signal, leaving the potential of textual feedback unexplored.
These methods also typically require numerous interactions with the environment to gather sufficient context, leading to significant computational overhead.
In this work, we propose a novel framework that takes a prompted large language model as an optimiser and treats the text-based prompt itself as a parameter.
By interacting with the environment to collect feedback, our proposed method constructs the updated textual prompt.
Our experimental results demonstrate that this method not only achieves superior performance but also automatically incorporates domain-specific knowledge, establishing a scientifically motivated, practical and efficient approach to prompting for future research.
Primary Area: transfer learning, meta learning, and lifelong learning
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.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 9516
Loading