LogicJitter: Let LLMs play Logic Games and they will Detect Misinformation
Keywords: llm, misinformation, rule based AI, toxicity
Abstract: In the face of the growing challenge of information overload online, the ability to accurately distinguish between genuine information and misinformation has become increasingly critical both from an individual and from a societal point of view. Methodologies for misinformation detection predominantly rely on supervised approaches, which depend heavily on large labeled datasets. However, these datasets are not only costly and time-consuming to produce, but they are also susceptible to issues such as labeling bias, time leakage, the inherent subjectivity of the task, and domain-specific limitations.
In this paper, we aim to overcome the aforementioned challenges by proposing a novel and cost-effective strategy to enhance the logical reasoning capabilities of Large Language Models (LLMs), thereby improving their ability to detect misinformation. Our approach, termed LogicJitter, employs a data augmentation technique during fine-tuning that generates both correct and incorrect statements within rule-based logic games. These games are designed to counteract well-known human cognitive biases and logical fallacies.
Hence, the primary contributions of this work include demonstrating the effectiveness of logical reasoning fine-tuning on LLMs and providing an open source package for the automatic generation of correct and incorrect logic-based training data, to ease reproducibility. Experimental results confirm this approach improves misinformation detection.
Primary Area: alignment, fairness, safety, privacy, and societal considerations
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: 6757
Loading