Memorisable Prompting: Preventing LLMs Forgetting False Positive Alarm
Keywords: Prompt-based task, Large language model, Memorisable Prompting for Data Annotation
Abstract: Large Language Models (LLMs) are widely recognized for their superior performance across various domains. However, their tendency to generate inaccurate or misleading responses presents significant challenges, particularly in the natural language domain. This issue underscores the need to enhance both the explainability and reliability of LLMs. While recent advancements in prompting have focused on leveraging in-context learning—such as providing step-by-step explanations—these approaches often overlook the critical importance of understanding the response dependency of LLMs on specific datasets. This understanding is crucial for interpreting their outputs and improving their consistency. Moreover, if we can capture and encode these response dependencies, we can integrate them into LLMs as memorized knowledge to mitigate false positive predictions over time. In this paper, we tackle this challenge by introducing the Memorizable Prompting (MP) paradigm, which enables LLMs to retain and utilize information from past responses. Specifically, our approach leverages hint samples—a small set of annotated examples—to learn the response dependencies, defined as the relationship between LLM outputs and the ground-truth annotations for a given dataset. This equips LLMs with the ability to recall past false positives and use that knowledge for self-correction in future predictions. We have evaluated our method on a diverse set of domain-specific datasets, demonstrating its effectiveness across large-scale benchmarks.
Primary Area: foundation or frontier models, including LLMs
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: 6711
Loading