Generalized Adversarial Learning--An Innovative Unsupervised Paradigm In LLM's Calibration
Keywords: Natural Language Processing, Adversarial Learning, Unsupervised Learning
Abstract: Recently, there has been a significant increase in the use of large-scale Question-Answering (QA) models. However, these models have started to reveal some limitations, such as generating incorrect information, which are holding back their progress. Most of the current calibration methods have their own problems, like the high cost of collecting fine-tuning data, limited interpretability and high intrusiveness, making them less suitable for wider use. To tackle this challenge, we introduce a new machine learning paradigm called "Generalized Adversarial Learning" (GAL) to improve the calibration of large QA models without the need for supervision. We explain the core principles and ideas behind GAL and present empirical evidence demonstrating its effectiveness, interpretability, and non-intrusive nature, achieving performance surpassing the state-of-the-art in some metrics even within the field of supervised learning.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation 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/2024/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: 5933
Loading