OOD Learner via In-Context Learning

ICLR 2026 Conference Submission22023 Authors

Keywords: OOD Detection, In-Context Learning

Abstract: Out-of-distribution (OOD) detectors are built on classification models to identify test samples that do not belong to any of their training classes. For classifiers based on pretrained vision-language models (VLMs), recent methods construct OOD detectors using text and few shot in-distribution (ID) images. In this work, we introduce a versatile framework for few-shot OOD detection through in-context learning (ICL). Instead of building an OOD detector for specific ID datasets, we propose a universal OOD Learner that can adapt to arbitrary ID datasets using few-shot texts and images as context, without the need for fine-tuning. Our method is implemented as an attention-based module and pretrained on pseudo-class data curated from large-scale text-image pairs. Experimental results demonstrate that our framework achieves state-of-the-art performance and efficiency in few-shot OOD detection.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 22023