GNN-as-Judge Transforms Low-Resource Learning for Graphs

Published on April 13, 2026

In the realm of text-attributed graphs (TAGs), large language models (LLMs) have traditionally excelled deep understanding of textual features. However, their performance faltered in low-resource settings, hampered of sufficient labeled data. This limitation stymied their potential to predict effectively in environments where intricate structural patterns emerged.

The introduction of GNN-as-Judge marks a pivotal shift in this landscape. This innovative framework tackles two critical issues: the challenge of generating trustworthy pseudo labels and the need to reduce label noise during model fine-tuning. Neural Networks (GNNs) with LLMs, GNN-as-Judge enhances learning efficiency even when labeled data is sparse.

Using a collaborative pseudo-labeling strategy, GNN-as-Judge identifies impactful unlabeled nodes through the lens of labeled ones. It then analyzes patterns of agreement and disagreement between LLMs and GNNs to refine label generation. A weakly-supervised fine-tuning algorithm complements this process, allowing for the effective use of informative pseudo labels while addressing noise concerns.

Experiments across multiple TAG datasets reveal that GNN-as-Judge outperforms existing methods, especially in low-resource scenarios. This development not only enhances the usability of LLMs in graph learning but also sets a new standard for addressing label scarcity and improving accuracy in complex environments.

Related News