Temporal Misinformation Detection: Simple Ways to Improve Temporal Generalization and Better Evaluate Language Models

ICLR 2026 Conference Submission17431 Authors

Keywords: Misinformation, Temporality, Temporal bias, Dataset

Abstract: Most of the current misinformation detectors display a lack of temporal generalization, despite the increasing reported scores in the literature. This can be attributed to classical machine learning evaluation protocols based on random splits. While widely adopted, these protocols often fail to reflect real-world model performance, a limitation that is particularly critical in misinformation detection, where temporal dynamics play a central role. In this paper, we present a comprehensive analysis of temporal biases across multiple misinformation datasets, with a specific focus on the temporal distribution of labels. We also introduce simple yet effective methods to improve performance in scenarios where temporal generalization is critical for NLP tasks. Our findings show that classical evaluation protocols tend to overestimate model performance in misinformation detection. To address this, we propose FC30, a new dataset, and introduce a general-purpose evaluation metric designed to better assess models under temporal shift and capture potential temporal bias.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 17431