Reproduce “Modality-free Graph In-context Alignment” 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
Wei Zhuo, Siqiang Luo
MF-GIA adapts graph encoders to new graphs from few labeled examples without converting encoded data to text or fine-tuning pretrained parameters. It computes a one-step gradient fingerprint from support examples, maps it to domain-conditioned feature and local-label transformations, and uses episodic training with DPAA. On BlogCatalog, it reaches 67.31% ± 2.60 versus Prodigy’s 53.44% ± 2.78 in 6-way, 5-shot evaluation.
Key insights
- "Gradient fingerprints," derived from single-step gradient updates, can effectively represent a graph's intrinsic domain characteristics, enabling modality-free domain embedding.
- Lightweight, domain-conditioned affine transformations (FiLM layers) allow for aligning heterogeneous pre-encoded features and indexed labels into a unified semantic space.
- Episodic meta-learning with Dual Prompt-Aware Attention (DPAA) is effective for training a graph model to perform in-context reasoning across diverse, unseen graph domains and tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.