LANO: Large Language Models as Active Annotation Agents for Open-World Node Classification

ICLR 2026 Conference Submission12921 Authors

Keywords: Node Classification, Open-world Graph Learning, Large Language Models

Abstract: Node classification is a fundamental task in graph learning. While Graph Neural Networks (GNNs) have achieved remarkable success in this area, their effectiveness relies heavily on large amounts of high-quality labels, which are costly to obtain. Moreover, GNNs are typically developed under a closed-world assumption, where all nodes belong to a fixed set of categories. In contrast, real-world graphs follow an open-world setting, where newly emerging nodes often stem from out-of-distribution (OOD) classes, making it challenging for GNNs to generalize. Motivated by the strong zero-shot reasoning and generalization ability of Large Language Models (LLMs), we propose LANO (LLMs as Active Annotation Agents for Open-World Node Classification). Our framework first aligns GNN representations with LLM token embeddings via instance-aware and feature-aware self-supervised learning, enabling LLMs to serve as zero-shot predictors for graph tasks. LANO then employs an influence- and uncertainty-driven strategy to select the most representative nodes and leverages LLMs for cost-effective pseudo-label generation. To suppress the spread of inaccurate labels and mitigate labeling bias, a soft feedback propagation mechanism disseminates bias-reduced pseudo labels to neighboring nodes with label decay mechanism, followed by iterative GNN optimization. Extensive experiments on multiple benchmarks demonstrate that LANO consistently outperforms popular baselines, showcasing the great potential of LLMs as active annotation agents for advancing open-world graph learning.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 12921