The paper introduces a novel post hoc out-of-distribution (OOD) detection method called NegLabel, which leverages a large number of negative labels from extensive corpus databases to enhance the detection of OOD samples. The method aims to improve the separability between in-distribution (ID) and OOD samples by selecting high-quality negative labels using the NegMining algorithm. These negative labels are chosen based on their semantic divergence from ID labels, ensuring that they have lower affinities with ID samples than OOD samples. The proposed NegLabel score combines the similarities of images with both ID and negative labels, effectively utilizing the text comprehension capabilities of vision-language models (VLMs). Extensive experiments demonstrate that NegLabel achieves state-of-the-art performance on various OOD detection benchmarks and generalizes well across multiple VLM architectures. Additionally, NegLabel shows strong robustness against diverse domain shifts. The key contributions of the paper include the introduction of massive negative labels, the development of the NegMining algorithm, and the theoretical analysis of negative labels in multi-label classification.The paper introduces a novel post hoc out-of-distribution (OOD) detection method called NegLabel, which leverages a large number of negative labels from extensive corpus databases to enhance the detection of OOD samples. The method aims to improve the separability between in-distribution (ID) and OOD samples by selecting high-quality negative labels using the NegMining algorithm. These negative labels are chosen based on their semantic divergence from ID labels, ensuring that they have lower affinities with ID samples than OOD samples. The proposed NegLabel score combines the similarities of images with both ID and negative labels, effectively utilizing the text comprehension capabilities of vision-language models (VLMs). Extensive experiments demonstrate that NegLabel achieves state-of-the-art performance on various OOD detection benchmarks and generalizes well across multiple VLM architectures. Additionally, NegLabel shows strong robustness against diverse domain shifts. The key contributions of the paper include the introduction of massive negative labels, the development of the NegMining algorithm, and the theoretical analysis of negative labels in multi-label classification.