GLIP-OOD: Zero-Shot Graph OOD Detection with Graph Foundation Model

TL;DR

GLIP-OOD advances zero-shot graph OOD detection by leveraging a Graph Foundation Model to align node representations with semantically rich label embeddings, and by using an LLM to generate pseudo-OOD labels from unlabeled data. In the ideal setting with all ID and OOD label names, the GFM achieves state-of-the-art OOD detection without node-level supervision; in practical scenarios with only ID labels, pseudo-OOD labels produced by the LLM substantially boost performance, approaching supervised baselines. Across four text-attributed graph benchmarks, GLIP-OOD demonstrates strong OOD discrimination and robustness to label-space configurations and scoring choices. This work highlights a scalable, label-efficient approach for open-world graph learning by fusing graph foundation models with language-model–driven label augmentation.

Abstract

Out-of-distribution (OOD) detection is critical for ensuring the safety and reliability of machine learning systems, particularly in dynamic and open-world environments. In the vision and text domains, zero-shot OOD detection - which requires no training on in-distribution (ID) data - has advanced significantly through the use of large-scale pretrained models, such as vision-language models (VLMs) and large language models (LLMs). However, zero-shot OOD detection in graph-structured data remains largely unexplored, primarily due to the challenges posed by complex relational structures and the absence of powerful, large-scale pretrained models for graphs. In this work, we take the first step toward enabling zero-shot graph OOD detection by leveraging a graph foundation model (GFM). Our experiments show that, when provided only with class label names for both ID and OOD categories, the GFM can effectively perform OOD detection - often surpassing existing "supervised" OOD detection methods that rely on extensive labeled node data. We further address the practical scenario in which OOD label names are not available in real-world settings by introducing GLIP-OOD, a framework that uses LLMs to generate semantically informative pseudo-OOD labels from unlabeled data. These generated OOD labels allow the GFM to better separate ID and OOD classes, facilitating more precise OOD detection - all without any labeled nodes (only ID label names). To our knowledge, this is the first approach to achieve node-level graph OOD detection in a fully zero-shot setting, and it attains performance comparable to state-of-the-art supervised methods on four benchmark text-attributed graph datasets.

Figures (3)

• Figure 1: Radar plot (AUROC and AUPR) on four text-attributed graph datasets. GLIP-OOD-R uses real OOD class names, while GLIP-OOD-L employs LLM-generated pseudo-OOD names. With no node-level supervision, GLIP-OOD-R exceeds supervised OOD baselines, and GLIP-OOD-L remains competitive even without real OOD labels.
• Figure 2: GLIP-OOD overview. Initially, the entire graph is unlabeled, and the labels of all nodes are unknown. GLIP-OOD leverages an LLM to generate semantically informative pseudo-OOD labels from unlabeled data, enabling the GFM to capture nuanced semantic boundaries between ID and OOD classes without requiring any labeled nodes.
• Figure 3: Visualization of the semantic label space. Each point corresponds to a class label name embedded by GLIP-OOD (green for ID labels, purple for real OOD labels, and blue for pseudo-OOD labels by the LLM). The pseudo-OOD labels lie outside the ID cluster but remain semantically close to real OOD labels, highlighting the effectiveness as proxies for enhancing OOD awareness.