Prompt Optimization Meets Subspace Representation Learning for Few-shot Out-of-Distribution Detection

TMLR Paper7593 Authors

Abstract: The reliability of artificial intelligence (AI) systems in open-world settings depends heavily on their ability to flag out-of-distribution (OOD) inputs unseen during training. Recent advances in large-scale vision-language models (VLMs) have enabled promising few-shot OOD detection frameworks using only a handful of in-distribution (ID) samples. However, existing prompt learning-based OOD methods largely overlook the geometry of the visual feature embeddings learned by VLMs whose structure is particularly informative for distinguishing ID from OOD data and holds rich representation capacity as they are pre-trained on millions of samples. To address this, we introduce a \textit{geometry-aware context optimization framework} that integrates subspace representation learning with prompt tuning. By projecting ID-relevant features into a subspace spanned by prompt vectors and simultaneously projecting ID-irrelevant components via orthogonal null-space projections, our approach strengthens the discriminative power of the learned prompt vectors, thereby leading to enhanced ID–OOD separability at test time. To enable an easy-to-handle, end-to-end learning under this framework, we design a geometry-regularized learning criterion that ensures strong OOD detection performance as well as high ID classification accuracy across settings. Moreover, the proposed framework can be seamlessly integrated with a wide range of existing context optimization methods, effectively complementing their softmax-based OOD detectors. Experiments on various real-world datasets showcase the effectiveness of our approach for reliable open-world AI systems.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Jaeho_Lee3

Submission Number: 7593