TriPrompt: Progressive Local Prompting for Few-shot Out-of-Distribution Detection

Few-shot out-of-distribution (OOD) detection is essential for deploying recognition models in open-world environments, where a reliable model should correctly recognize in-distribution (ID) samples while rejecting samples from unseen categories. Recent CLIP-based methods have shown promising few-shot OOD detection ability by exploiting transferable vision--language representations. However, many methods rely mainly on global image features or emphasize only the most class-discriminative local cues, making them less effective for OOD detection where ID and OOD samples share visually similar local patterns. Although local prompting introduces patch-level reasoning, existing methods often treat local evidence as a binary distinction between ID-related and non-ID cues, leaving the ambiguous transition region between reliable ID evidence and clear non-ID evidence insufficiently modeled. To address this issue, we propose ThriPrompt, a progressive local prompting framework for few-shot OOD detection. TriPrompt decomposes patch-level evidence into three complementary components within a unified local feature space: Core-ID cues that provide reliable class-defining evidence, Far-ID cues that capture ambiguous transition-region evidence, and OOD cues learned from weak non-ID proxies mined directly from patch features. This patch-native design enables suppression-aware local reasoning without requiring additional OOD data or repeated multi-crop feature extraction. At inference time, \modelname{} combines global CLIP confidence with Core-ID, Far-ID, and OOD local responses to reduce overconfident ID assignment on hard OOD samples. Experiments on ImageNet-1K OOD benchmarks show that \modelname{} consistently improves few-shot OOD detection performance and remains compatible with mainstream CLIP-based scoring methods. Our code is provided in the supplementary material.