This paper investigates the problem of shortcut learning in safety guardrails for large language models (LLMs). It reveals that current safeguard models often rely excessively on superficial cues, such as specific keywords that are spuriously correlated with training labels, rather than genuinely understanding the input’s semantics or intent. As a result, their performance degrades significantly when there is a shift in keyword distribution. The paper also examines the impact of reducing shortcut reliance, showing that merely minimizing shortcut influence is insufficient. To build robust safeguard models, it is equally crucial to promote the use of intended features.

Entity disambiguation (ED) is crucial in natural language processing (NLP) for tasks such as question-answering and information extraction. A major challenge in ED is handling overshadowed entities—uncommon entities sharing mention surfaces with common entities. The current approach to enhance performance on these entities involves reasoning over facts in a knowledge base (KB), increasing computational overhead during inference. We argue that the ED performance on overshadowed entities can be enhanced during training by addressing shortcut learning, which does not add computational overhead at inference. We propose a simple yet effective debiasing technique to prevent models from shortcut learning during training. Experiments on a range of ED datasets show that our method achieves state-of-the-art performance without compromising inference speed. Our findings suggest a new research direction for improving entity disambiguation via shortcut learning mitigation.

Dense retrieval is a basic building block of information retrieval applications. One of the main challenges of dense retrieval in real-world settings is the handling of queries containing misspelled words. A popular approach for handling misspelled queries is minimizing the representations discrepancy between misspelled queries and their pristine ones. Unlike the existing approaches, which only focus on the alignment between misspelled and pristine queries, our method also improves the contrast between each misspelled query and its surrounding queries. To assess the effectiveness of our proposed method, we compare it against the existing competitors using two benchmark datasets and two base encoders. Our method outperforms the competitors in all cases with misspelled queries. Our code and models are available at https://github.com/panuthept/DST-DenseRetrieval.