Reproduce “Test-Time Training on Graphs with Large Language Models (LLMs)” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiaxin Zhang, Yiqi Wang, Xihong Yang, Siwei Wang, Yu Feng, Yu Shi, Ruicaho Ren, En Zhu and 1 more
Researchers at the National University of Defense Technology introduced LLMTTT, a framework that leverages Large Language Models to generate pseudo-labels for a selected subset of test nodes, thereby enhancing Graph Neural Network performance on out-of-distribution graphs. This method consistently achieved superior generalization across various datasets by transforming the unsupervised test-time adaptation problem into a more manageable semi-supervised learning task.
Key insights
- Leveraging LLMs to provide a small set of pseudo-labels effectively transforms the challenging unsupervised TTT problem into a more tractable semi-supervised learning scenario at test time.
- Combining model uncertainty with data distribution characteristics in active learning enables the selection of high-impact nodes for LLM annotation, significantly improving adaptation.
- Even a limited number of high-quality, strategically chosen pseudo-labels for test nodes can substantially reduce the generalization error bound on the test domain, as supported by theoretical analysis.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.